Blahalla

Frame Security Emerges From Stealth With $50M for Awareness and Training Platform

www.securityweek.com · 5 months ago

Frame Security emerged from stealth on May 11, 2026 with 50 million dollars in funding from Team8, Index Ventures, Picture Capital, Elad Gil, Cerca Partners, and Tesonet. The company, founded by Tal Shlomo as CEO and Sharon Shmueli as CTO, has built an AI powered human risk management platform covering the full security awareness lifecycle from simulated attacks to employee training to threat triage. Its simulation module generates personalized phishing, voice, and video deepfake scenarios tailored to employee roles and communication patterns, while a continuous human risk scoring engine aggregates signals from simulations and employee behavior to provide real time risk visibility at individual, team, and organizational levels. A phishing and threat triage module lets employees report suspicious messages across any communication channel and uses AI to analyze and score them in real time, and the platform is already used by startups and Fortune 500 companies.
How I Defeat Passkeys Nearly Every Time in Phishing Assessments

infosecwriteups.com · 5 months ago

NULL
Entries now open for the 2026 CSO30 Australia Awards

www.csoonline.com · 5 months ago

Nominations are now open for the 2026 CSO30 Australia Awards celebrating the country’s most effective and influential cybersecurity leaders, with entries closing on 26 June 2026. The awards will be presented alongside the CIO50 awards at a flagship event on 22 September 2026 in Sydney. Nominees will be assessed across two core pillars, business value and leadership, based on cybersecurity initiatives delivered over the past two years and reviewed confidentially by an independent panel of judges. The program also recognises emerging talent through the Next CISO Award and is open to senior cybersecurity leaders responsible for defining and executing security strategy within Australian organisations.
Can the U.S. Rein in Prediction Markets? + Joanna Stern on Her Year of A.I. Experiments + Our Producer Goes to Attention School

www.nytimes.com · 5 months ago

false
1.8 Billion Gmail Users May Want to Check This AI Privacy Setting

www.techrepublic.com · 5 months ago

Google announced on May 7 new Gmail AI personalization updates that let Gemini analyze Gmail and Google Drive content to provide topic contextualization, contextual replies, smart searches, and tone and style personalization for drafting. Users can disable these smart personalization features by opening Gmail settings and turning off Smart features and personalization, but doing so may stop receiving new AI upgrades and does not fully prevent automated processing or Google access to the account data required to run the service. The changes are likely enabled by default, have prompted public warnings and a recent class action over automatic enabling of Smart Features, and users concerned about sensitive information are advised to review their smart settings.
Hackers Used AI to Develop First Known Zero-Day 2FA Bypass for Mass Exploitation

thehackernews.com · 5 months ago

There is an unknown threat actor using a zero day exploit likely developed with an AI system to bypass two factor authentication in a popular open source web based system administration tool. The exploit was implemented as a Python script that shows hallmarks of large language model generated code and requires valid user credentials, exploiting a semantic logic flaw caused by a hard coded trust assumption. Google worked with the impacted vendor to disclose and fix the flaw, disabled related malicious infrastructure, and linked the activity to an Android backdoor called PromptSpy that can capture biometrics, manipulate the user interface, and update command and control components dynamically. Google also reported that multiple criminal and nation linked groups are using AI, shadow APIs, and specialized tools to accelerate vulnerability discovery, exploit generation, and autonomous malware operations.
Instructure confirms hackers used Canvas flaw to deface portals

www.bleepingcomputer.com · 5 months ago

There are multiple cross-site scripting vulnerabilities that allowed attackers to obtain authenticated admin sessions and modify Canvas login portals to display an extortion message. The initial breach was discovered on April 29, 2026, with Instructure confirming stolen data and ShinyHunters claiming more than 3.6 terabytes of exfiltrated data and about 275 million records affecting roughly 8,809 educational organizations; the actor returned on May 7, 2026 to deface portals and demand ransom. Instructure temporarily shut down Free-for-Teacher accounts, took Canvas offline to apply safeguards, restored Canvas on May 9, 2026, and says the portal defacements did not themselves expose additional data though the initial exfiltration likely included usernames, emails, course and enrollment information.
'Dirty Frag' Exploit Poised to Blow Up on Enterprise Linux Distros

www.darkreading.com · 5 months ago

A public exploit is available for a nine year old Linux kernel privilege escalation vulnerability called Dirty Frag that chains two flaws and affects a wide range of enterprise Linux distributions including Ubuntu, Red Hat Enterprise Linux, CentOS Stream, AlmaLinux, openSUSE Tumbleweed, and Fedora. The two issues are tracked as CVE-2026-43284 and CVE-2026-43500, each assigned a 7.8 CVSS score and an Important severity rating by Red Hat, and a proof of concept has been published with limited in the wild activity observed that may indicate exploitation. Upstream Linux released a patch for CVE-2026-43284 while fixes for CVE-2026-43500 are not yet available and major vendors are expediting or preparing kernel updates and livepatches. Until patches are applied, recommended mitigations include disabling unused rxrpc modules, assessing whether esp4 esp6 and related xfrm IPsec functionality can be safely disabled, restricting local shell access, hardening container workloads, enforcing SELinux, running workloads as non root, and increasing monitoring for abnormal privilege escalation activity.
Lyrie.ai Joins First Batch of Anthropic’s Cyber Verification Program

gbhackers.com · 5 months ago

OTT Cybersecurity LLC’s Lyrie dot ai has been accepted into Anthropic’s Cyber Verification Program and concurrently released the Agent Trust Protocol, an open cryptographic standard that specifies agent identity, scope, attestation, delegation, and revocation. The Agent Trust Protocol is open and royalty free, slated for submission to the Internet Engineering Task Force, and has a reference implementation published under an MIT license to enable real time verification that an AI agent is who it claims to be, what it is authorized to do, and whether it or its instructions have been tampered with. Lyrie is presented as a unified offensive and defensive security platform for the agentic AI era with features including an autonomous seven phase penetration test command, GPU accelerated red teaming, OWASP Agentic Security Initiative coverage, autonomous zero day discovery workflows, and hardware flexible deployment.
Build Application Firewalls Aim to Stop the Next Supply Chain Attack

www.securityweek.com · 5 months ago

There are supply chain attacks that compromise development toolchains, exemplified by the 2020 SolarWinds breach and a string of March 2026 incidents that hijacked an npm maintainer to publish malicious Axios versions and compromised tools including Trivy and LiteLLM, ultimately delivering a remote access trojan and causing the European Commission to lose 300 gigabytes of data. Standard CI slash CD scanners often miss stealthy malicious changes or unknown zero day vulnerabilities because malicious behavior can appear benign in code or manifests. Inspecting runtime behavior with a build application firewall that enforces policy during builds, performs deep packet inspection, flags unexpected actions with AI explanations, and produces a full accurate software bill of materials is proposed as a way to detect exfiltration and reduce supply chain risk.
Why Changing Passwords Doesn’t End an Active Directory Breach

www.bleepingcomputer.com · 5 months ago

Password resets do not always evict attackers from Active Directory or hybrid Entra ID environments because Windows can retain cached password hashes for offline logon, Kerberos tickets remain valid for their set lifetime, and synchronization delays can leave old hashes usable. Attackers exploit cached hashes with techniques like pass the hash, maintain access via active Kerberos sessions or forged tickets including Golden and Silver Ticket attacks, and rely on long‑lived service account credentials or modified ACLs to preserve access. Mitigations include terminating active sessions and purging Kerberos tickets by forcing logoffs or reboots, rotating and in some cases resetting critical credentials such as the KRBTGT account, rotating service account passwords, clearing cached credentials as systems reconnect, forcing AD to Entra ID synchronization, and auditing group memberships and ACLs to remove hidden backdoors.
fsnotify Maintainer Access Change Sparks Supply Chain Security Concerns

gbhackers.com · 5 months ago

A dispute over maintainer access in the fsnotify Go library triggered supply chain concerns after contributors were removed from the project’s GitHub organization and recent releases came under scrutiny. No evidence indicates any version of fsnotify has been compromised, but the timing of new releases after a long quiet period, revoked access, and unclear review processes resembled early warning signs of supply chain attacks. Downstream projects including Kubernetes began evaluating the library and potential alternatives, and a fork named gofsnotify slash fsnotify emerged as a monitored fallback. The incident highlights the need for transparent governance, clearly defined release authority, and multiple active maintainers to preserve trust in critical open source components.
Lyrie.ai Joins First Batch of Anthropic’s Cyber Verification Program

www.csoonline.com · 5 months ago

There are two developments from OTT Cybersecurity LLC: Lyrie.ai has been accepted into Anthropic's Cyber Verification Program and the company publicly released the Agent Trust Protocol, an open cryptographic standard for verifying autonomous AI agents. The Agent Trust Protocol provides real time verification of an agent's identity, authorized scope, attestation of tampering, delegation, and revocation, is open and royalty free, and is slated for submission to the Internet Engineering Task Force. Lyrie is positioned as an integrated offensive and defensive cybersecurity platform for the AI agent era, offering autonomous penetration testing workflows, GPU powered red teaming, binary analysis for zero day discovery, alignment with the OWASP Agentic Security Initiative taxonomy, and a reference implementation released under an MIT license.
Google Detects First AI-Generated Zero-Day Exploit

www.securityweek.com · 5 months ago

Google detected a zero-day exploit believed to have been developed using artificial intelligence, marking the first known AI generated zero-day. The exploit was implemented as a Python script and engineered to bypass two factor authentication on an open source web based system administration tool, and Google worked with the affected vendor to prevent mass exploitation while the threat actor and targeted tool remain unnamed. Google noted the script included textbook style docstrings and hallucinated metadata consistent with large language model output, and said Chinese and North Korean state linked actors have increasingly used AI driven tools to accelerate vulnerability discovery and exploit development.
Google: Hackers used AI to develop zero-day exploit for web admin tool

www.bleepingcomputer.com · 5 months ago

Researchers at Google's Threat Intelligence Group say a zero-day exploit targeting a popular open-source web-based system administration tool was likely generated using an AI model and could bypass the tool's two-factor authentication. Google says the exploit's Python code contains features characteristic of large language model output, including educational docstrings, a hallucinated CVSS score, and a structured Pythonic style, giving researchers high confidence the actor used AI to find and weaponize the flaw. The attack was disrupted before mass exploitation, and Google rules out its Gemini model being involved while noting Chinese, North Korean, and other linked groups have increasingly relied on AI for vulnerability discovery and exploit development. Google warns threat actors are industrializing access to premium AI models through automated account creation, proxy relays, and account pooling, and are using AI for operations such as voice cloning and automated malware components.
Google Says Criminal Hackers Used A.I. to Find a Major Software Flaw

www.nytimes.com · 5 months ago

A prominent criminal group used artificial intelligence to discover a previously unknown software vulnerability and to develop a zero day exploit that targeted two factor authentication before Google disrupted the operation. Google's Threat Intelligence Group said this marks the first observed instance of attackers using AI to find a new bug and warned that adversaries, including state linked actors, are integrating AI across the attack lifecycle. Experts said the incident is a taste of what is to come because AI can accelerate vulnerability discovery, exploit development, malware obfuscation, and automated attack orchestration, prompting calls for stronger defenses and pre release model vetting.
Hackers Use AI for Exploit Development, Attack Automation

www.darkreading.com · 5 months ago

There is increasing use of large language models and agentic tools by threat actors to assist vulnerability research, develop exploits, and orchestrate multi stage attacks. Google's Threat Intelligence Group identified a probable AI developed zero day implemented in a Python script that bypasses two factor authentication on a popular open source web based system administration tool and noted characteristic AI artifacts in the exploit such as educational docstrings and a hallucinated CVSS score. State linked and criminal actors are prompting models to recursively analyze CVEs, training on legacy vulnerability repositories, and using agentic frameworks and malicious Android backdoors that leverage AI to navigate user interfaces, capture biometrics, and automate persistence, signaling a shift toward AI driven autonomous attack frameworks with reduced human oversight.
Google discovers weaponized zero-day exploits created with AI

www.csoonline.com · 5 months ago

Google Threat Intelligence Group reported evidence that a cybercriminal group used artificial intelligence to create a weaponized zero day exploit that bypasses two factor authentication in a popular open source web based system administration tool. The exploit was a Python script that required credentials to run and exploited a hardcoded trust assumption by the tool’s developers, and GTIG disclosed the vulnerability to the vendor which likely hindered mass exploitation. Researchers identified artifacts consistent with large language model generation, including educational strings, a hallucinated CVSS score, and textbook style Python elements, supporting the conclusion that an AI model both discovered the vulnerability and authored the exploit. GTIG also observed other threat actors experimenting with frontier models and agentic tools to analyze firmware and protocols, indicating AI is increasingly able to surface high level logic errors and accelerate exploit development.
⚡ Weekly Recap: Linux Rootkit, macOS Crypto Stealer, WebSocket Skimmers and More

thehackernews.com · 5 months ago

There are active exploitations of Ivanti Endpoint Manager Mobile tracked as CVE-2026-6973 and of Palo Alto Networks PAN-OS tracked as CVE-2026-0300, the latter allowing unauthenticated root remote code execution with patches expected to begin May 13, 2026. There is a new modular Linux remote access trojan called Quasar Linux or QLNX that combines a kernel-level rootkit, PAM authentication backdoors, persistence mechanisms, and a peer-to-peer mesh to create resilient interconnected infections and harvest credentials. Multiple supply chain and cloud-focused campaigns were reported, including a DAEMON Tools compromise that deployed a QUIC RAT, a JDownloader compromise on May 6, 2026 that distributed malicious installers, and a PCPJack campaign observed replacing TeamPCP tooling to steal cloud and developer credentials. There are numerous trending CVEs to prioritize for patching, notably CVE-2026-6973 and CVE-2026-0300, along with other high-impact entries such as CVE-2026-4670, CVE-2026-5174, CVE-2026-43284, and CVE-2026-43500.
cPanel and WHM Servers Targeted in Attacks Exploiting CVE-2026-41940

gbhackers.com · 5 months ago

There is a critical authentication bypass vulnerability in cPanel and WHM servers identified as CVE-2026-41940 with a CVSS score of 9.8 that permits unauthenticated remote attackers to obtain full administrator privileges. A long running, highly skilled syndicate known as MrRot13 is actively exploiting the flaw since public disclosure in late April 2026 to deploy Go based payload injectors, implant persistent backdoors and gain comprehensive remote control. Attackers inject malicious JavaScript into cPanel login templates to capture credentials and session data, exfiltrate harvested information via ROT13 obfuscated domains or a Telegram channel, and have been linked to incidents that exfiltrated over four gigabytes from Southeast Asian government and military networks. Observed indicators include multiple MD5 hashes for malicious binaries and defanged command and control domains to assist detection and incident response.
Webinar this week: Prevention alone is not enough against modern attacks

www.bleepingcomputer.com · 5 months ago

Modern cyberattacks now combine AI generated phishing, business email compromise, ransomware, and abuse of legitimate SaaS and trusted infrastructure to gain access and disrupt operations. A webinar on Thursday, May 14, 2026 at 2:00 PM Eastern will explain why prevention alone is insufficient and why organizations must couple prevention and detection with reliable backups and rapid recovery to limit downtime and data loss. The session will cover AI driven phishing and brand impersonation, how attackers leverage trusted platforms to bypass defenses, common failure points after initial compromise, and the critical role of SaaS backups and business continuity and disaster recovery planning.
Malicious Hugging Face model masquerading as OpenAI release hits 244K downloads

www.csoonline.com · 5 months ago

A malicious Hugging Face repository impersonating OpenAI's Privacy Filter, named Open-OSS privacy filter, reached the platform's top trending position with about 244,000 downloads and 667 likes in under 18 hours and contained a loader script that fetched and executed credential-stealing malware on Windows systems. The loader disabled SSL verification, decoded a base64 encoded pointer to a public JSON hosting service to retrieve remote commands which were executed via PowerShell, downloaded an additional batch file, created a scheduled task that mimicked a Microsoft Edge update for persistence, and deployed a Rust based infostealer that stole data from Chromium and Firefox browsers, Discord storage, cryptocurrency wallets, FileZilla configurations, and other host information while attempting to disable Windows antimalware interfaces and evade sandbox analysis. Researchers found multiple related repositories and shared infrastructure tied to broader open-source supply-chain campaigns, warned that traditional software composition analysis tools are ill suited to detect malicious AI artifacts, urged affected users to treat impacted systems as fully compromised and reimage, block indicators of compromise, rotate credentials, invalidate active sessions, and hunt historical network connections, and reported that the repository was removed from the platform for violating terms of service.
Lyrie.ai Deploys Real-Time Zero-Day Tracking Across Global Enterprise Infrastructure

www.csoonline.com · 5 months ago

OTT Cybersecurity LLC announced deployment of a real time zero day tracking and disclosure system, acceptance into Anthropic's Cyber Verification Program, and the public release of the Agent Trust Protocol. Lyrie's autonomous threat intelligence engine continuously monitors global infrastructure, open source repositories, API surfaces, and agent to agent communication to identify zero day vulnerabilities as they emerge and can generate disclosure packages with proof of concept analysis, impact assessment, and remediation guidance delivered to affected organizations within hours. Anthropic accepted the company into its Cyber Verification Program to verify legitimate dual use cybersecurity operators and to permit vulnerability research and offensive tooling work subject to applicable safety policies. The Agent Trust Protocol is an open royalty free cryptographic standard and reference implementation that enables real time verification of AI agent identity, scope, attestation, delegation, and revocation and is slated for submission to the Internet Engineering Task Force.
Python Infostealer Hides in GitHub Releases to Bypass Detection

gbhackers.com · 5 months ago

A stealthy Python based infostealer campaign abuses GitHub Releases to host payloads and maintain long term low visibility access, using humanitarian themed phishing lures targeting Russian speaking users in an operation tracked as Operation HumanitarianBait. Infection begins with a RAR archive containing a large malicious Windows shortcut that hides self obfuscated Unicode decoded and executed by PowerShell, presents a decoy Russian language PDF, then downloads a PE less Python payload and supporting components from GitHub Releases to bootstrap a self contained Python environment under the user AppData WindowsHelper directory. The implant persists via VBScript launchers and a scheduled task named WindowsHelper, uses PyArmor to protect the main module, harvests browser credentials by extracting AES GCM master keys and supporting DPAPI and newer Chrome formats, logs keystrokes and clipboard, captures screenshots, collects Telegram session data, scans for documents and cryptocurrency keys, can install remote access tools stealthily, and exfiltrates data to a VPS hosted command and control server running an nginx fronted Flask dashboard. Security teams are urged to monitor for suspicious LNK files and GitHub Releases download activity, scheduled tasks named WindowsHelper, and Python or script execution from AppData paths.
PHP SOAP Extension Flaw Could Let Attackers Execute Code Remotely

gbhackers.com · 5 months ago

There are multiple PHP vulnerabilities affecting the SOAP extension and core modules, including a high severity use after free flaw identified as CVE-2026-6722 that can enable remote code execution. Additional issues include CVE-2026-7261 a moderate use after free tied to SoapServer session persistence, CVE-2026-7262 a NULL pointer dereference in the Apache Map decoder that can cause denial of service, CVE-2026-7258 an out of bounds read in the urldecode function, and CVE-2026-6104 a global buffer over read in mbconvertencoding that can lead to information disclosure. The flaws affect PHP releases before versions 8.2.31 8.3.31 8.4.21 and 8.5.6 and have been fixed in patched releases so administrators are urged to upgrade to 8.2.31 8.3.31 8.4.21 or 8.5.6 immediately.
New ‘Dirty Frag’ exploit targets Linux kernel for root access

www.csoonline.com · 5 months ago

There is a newly disclosed Linux privilege escalation issue called Dirty Frag that abuses fragmented page cache handling to overwrite protected data and enable local escalation to root, and researchers report active exploitation in the wild. The attack chain combines flaws in the IPsec Encapsulating Security Payload subsystem and the RxRPC networking protocol tracked as CVE-2026-43284 and CVE-2026-43500 and is related to earlier page cache attacks including Dirty Pipe CVE-2022-0847 and Copy Fail CVE-2026-31431. Observed use cases include post compromise escalation after SSH access, web shell execution, container escape, or compromise of a low privileged account, with successful exploitation able to grant broad control of the affected host. Because patches are rolling out unevenly and one fix remains pending, researchers recommend immediate mitigations such as disabling esp4 esp6 and rxrpc kernel modules if not required, reducing unnecessary local shell access, monitoring for abnormal privilege escalation, and tightening container runtime controls while noting that mitigation may not undo prior successful exploitation.
Palantir’s access to identifiable NHS England patient data is ‘dangerous’, MPs say

www.theguardian.com · 5 months ago

NHS England has allowed Palantir staff and other contractors to access identifiable patient information in its federated data platform, including parts described in internal briefings as having unlimited access for non NHS England staff, and in some cases the data was accessed before it was pseudonymised. The arrangement follows a 330 million pounds contract to build the platform and deploy AI to integrate scattered health datasets and has drawn warnings from MPs, patient groups, and campaigners that the decision is dangerous and will undermine public confidence in data privacy. NHS England says external engineers must have government security clearance and that access is logged and approved at director level, while Palantir says it acts as a data processor with granular controls that prevent removal or misuse of patient records.
Cyber Espionage Group Targets Aviation Firms to Steal Map Data

www.darkreading.com · 5 months ago

A cyber espionage group known as HeartlessSoul has targeted aerospace and drone operators to exfiltrate geospatial information including GIS files, terrain models, and GPS data to build detailed operational awareness. The group used tailored phishing and malvertising, created fake domains and installer sites, and even uploaded a malicious archive to a legitimate download service to deliver a JavaScript remote access trojan, PowerShell scripts, fileless execution techniques, and malicious LNK files. Compromised data types include GPS logs, GIS shapefiles, digital relief files, proprietary mapping files, and other documents, and some LNK files exploited a Windows shortcut vulnerability tracked as ZDI-CAN-25373. Security guidance recommends hunting for indicators, protecting GIS and flight‑planning systems with identity‑bound access and egress monitoring, segmenting engineering networks from business networks, and applying zero-trust protections to workstations that handle crown-jewel mapping data.
Skoda Data Breach Hits Online Shop Customers

www.securityweek.com · 5 months ago

There was a data breach affecting Skoda online shop customers that was discovered through technical security monitoring and caused by a vulnerability in the portal software. Attackers accessed customer names, addresses, email addresses, phone numbers, order details, user account information, and password hashes, while no credit card data was compromised because payments are handled by third party providers, and Skoda cannot determine if or to what extent data was exfiltrated or how many people were affected. Skoda took the shop offline, patched the vulnerability, retained external forensics experts, notified authorities, and advised customers to watch for phishing, avoid clicking suspicious links, and change reused passwords.
Your Purple Team Isn't Purple — It's Just Red and Blue in the Same Room

thehackernews.com · 5 months ago

Purple teaming aims to create a continuous loop where red finds attacker paths and blue validates detections, but system workflows and messy handoffs mean teams still often behave as separate red and blue functions. Three operational barriers are human friction, fragmented orchestration of teams and tools, and the inability to keep pace with AI-powered attackers; exploitation windows shrank from a mean of 56 days in 2024 to about 23 days in 2025 and roughly 10 hours so far in 2026 across more than three thousand CVE-exploit pairs, while AI-assisted attackers can compromise systems in about 73 seconds and defenders commonly take at least 24 hours to deploy fixes. Autonomous purple teaming uses automated penetration testing, breach and attack simulation, and chains of AI agents to turn red findings into blue tests, automate low-risk fixes, and maintain an auditable machine-speed loop with human review when needed.
Cloudflare Lays Off 1,100 Employees in AI-Driven Restructuring

www.securityweek.com · 5 months ago

Cloudflare is cutting more than 1,100 jobs worldwide as part of an AI driven restructuring after internal AI usage rose by over 600 percent in the prior three months. Affected employees will receive severance equal to their full base salary through the end of 2026, US healthcare coverage through December 31, 2026, accelerated equity vesting through August 15, 2026, and waived or pro rated vesting for those who have not reached a one year cliff. The company beat revenue and earnings forecasts for the first quarter of 2026, but its shares tumbled more than 20 percent following the layoff announcement and softer near term guidance.
LLMs and Text-in-Text Steganography

www.schneier.com · 5 months ago

Large language models can hide meaningful messages inside otherwise ordinary coherent text, enabling text-in-text steganography. A demonstrated protocol encodes and decodes messages of the same length using even modest eight billion parameter models, and can run locally on a laptop in seconds, creating covert channels that can bypass content filters. This decoupling of text from authorial intent raises urgent trust, safety, and content moderation challenges and points to the need for detection tools and policy responses.
Fake Claude Campaign Uses PlugX-Style DLL Sideloading Chain

gbhackers.com · 5 months ago

Hackers are abusing a fake Claude AI download site to deliver a PlugX-style DLL sideloading chain that drops NOVupdate.exe NOVupdate.exe.dat and avk.dll into the Startup folder and leverages a signed G DATA updater to load a malicious avk.dll. The chain decrypts shellcode which uses DonutLoader to execute a previously undocumented Windows backdoor named Beagle that can execute arbitrary commands upload and download files list and remove directories rename files and uninstall itself. Beagle communicates with command and control over TCP port 443 and UDP port 8080 using AES with a hardcoded key and per-packet IVs and the campaign uses Cloudflare fronting and specific IP infrastructure. Organizations are advised to obtain AI tool downloads only from official vendor domains and to detect signed G DATA executables loading nonstandard avk.dll files encrypted outbound traffic to the campaign domains and in-memory DonutLoader activity.
SailPoint Discloses GitHub Repository Hack

www.securityweek.com · 5 months ago

On April 20, 2026, unauthorized access to a subset of SailPoint's GitHub repositories was detected and immediately contained. The repositories were compromised via a vulnerability in a third party application that has since been addressed. An investigation conducted with an external cybersecurity firm found no evidence that customer data in SailPoint's production or staging environments was accessed or that services were interrupted. SailPoint directly notified customers whose information was stored in the accessed repositories, advised generally that no further action is required, did not name a threat actor and did not disclose additional details making any link to recent supply chain attacks unclear.
Microsoft 365 Copilot Flaws Could Let Attackers Access Sensitive Data

gbhackers.com · 5 months ago

There are three critical information disclosure vulnerabilities affecting Microsoft 365 Copilot and Copilot Chat in Microsoft Edge, disclosed May 7, 2026, that could allow attackers to bypass security boundaries and access sensitive enterprise data. The flaws, tracked as CVE-2026-26129, CVE-2026-26164, and CVE-2026-33111, each have CVSS 3.1 base scores of 7.5, require no privileges or user interaction, and exploit improper input handling and injection weaknesses in the Copilot processing pipeline and the Edge chat context. Microsoft has deployed backend fixes to the fully managed Copilot service and reports that no customer action is required to mitigate these information disclosure risks.
Trending Hugging Face Repo With 200K Downloads Spreads Windows Malware

gbhackers.com · 5 months ago

A malicious Hugging Face repository named Open-OSS/privacy-filter abused the platform’s trust and trending algorithm to appear legitimate, briefly reaching the number one trending position with about 244,000 downloads and hundreds of likes before removal. The repository typosquatted a legitimate Privacy Filter release, instructed users to run start dot bat or loader dot py on Windows, and the loader disabled SSL verification, fetched execution commands from a remote JSON paste, and invoked a hidden PowerShell one liner that downloaded an update script which elevated privileges, fetched a final payload, added Microsoft Defender exclusions, and scheduled a one shot task to run the payload as system. The final roughly one megabyte Rust based infostealer contains anti analysis and virtual machine detection, attempts to disable AMSI and ETW, harvests browser credentials, session tokens, cryptocurrency wallet data, FTP SSH and VPN files, and exfiltrates compressed data over HTTPS to a command and control server, so any Windows system that executed the repository files should be treated as fully compromised with full reimaging and credential rotation recommended.
Sandboxie Escape Flaw Could Let Attackers Gain SYSTEM-Level Privileges

gbhackers.com · 5 months ago

There are critical sandbox escape vulnerabilities in Sandboxie and Sandboxie-Plus that allow attackers to bypass isolation and obtain full SYSTEM level privileges. The flaws include CVE-2026-34459, a stack based buffer overflow in the SbieSvc GetRawInputDeviceInfoSlave handler that enables an ASLR bypass and return oriented programming chain, CVE-2026-34458, a configuration file injection via unsanitized carriage return and line feed handling, CVE-2026-32603, a local denial of service that triggers a kernel crash, and CVE-2026-34527, a hashing implementation error that reduces stored password entropy. These vulnerabilities affect Sandboxie versions 1.17.2 and earlier and can be chained to escalate privileges even against hardened sandbox configurations. The issues have been fixed in subsequent builds and users are urged to update to version 1.17.5 or uninstall vulnerable installs and then install the patched release.
AI security is repeating endpoint security’s biggest mistake

www.csoonline.com · 5 months ago

There is déjà vu in security as many teams treat AI security like a posture problem, relying on model cards, software bills of materials, input and output filters, prompt injection guardrails and access controls while the AI surface rapidly expands. That posture first approach is brittle because open source models deployed without review, third party model APIs embedded in SaaS, autonomous agents with broad access and retrieval augmented generation pipelines that touch sensitive data create shadow AI and dynamic opaque attack surfaces posture alone cannot secure. Behavioral detection that logs and analyzes action sequences such as unusual data access patterns, prompt injection artifacts, unexpected tool calls and token velocity anomalies provides the signal SOC teams need to triage incidents rather than audit findings. Organizations should retain posture controls but start recording AI behavior now, prioritize highest agency surfaces, think in sequences not single events and route behavioral telemetry to the SOC for detection and response.
Checkmarx Jenkins AST Plugin Compromised in Supply Chain Attack

www.securityweek.com · 5 months ago

A maliciously modified version of the Checkmarx Jenkins AST plugin was published to the Jenkins Marketplace as part of a supply chain attack. Checkmarx warned users and advised ensuring they are running version 2.0.13-829.vc72453fa1c16 published in December 2025, and released two new plugin versions over the weekend with the latest now available on GitHub and the Jenkins Marketplace. The compromise is tied to the Trivy supply chain attack that allowed the TeamPCP group to access Checkmarx repositories in late March, followed by additional malicious artifacts and alleged data leaked by the Lapsus extortion group.
PoC Exploit Released for Android Zero-Click Flaw Enabling Remote Shell Access

gbhackers.com · 5 months ago

A public proof of concept is circulating for CVE-2026-0073, a critical no-interaction remote code execution flaw in Android's adbd that can grant shell access when wireless ADB is reachable on a network. The weakness stems from a logic error in adbdtlsverifycert that can cause a crafted TLS client certificate to be treated as trusted, bypassing mutual authentication and allowing an unauthorized ADB host to open a shell session. Defenders are urged to apply the May 2026 Android security update, disable wireless debugging where not needed, and treat any LAN or internet reachable ADB service as a sensitive attack surface because public exploit references increase the likelihood of opportunistic scanning and targeted abuse.
TrickMo Android banker adopts TON blockchain for covert comms

www.bleepingcomputer.com · 5 months ago

A new variant of the TrickMo Android banking malware is being delivered in campaigns across Europe and targets banking credentials and cryptocurrency wallets in France, Italy, and Austria. It uses The Open Network TON overlay for covert command and control by resolving operator endpoints as.adnl identities through an embedded local TON proxy on infected devices, which hides IP addresses and makes traditional domain takedowns and network edge detection ineffective. The malware retains a modular two stage design and adds network diagnostic and proxying commands including curl, dnsLookup, ping, telnet, traceroute, SSH tunneling, remote and local port forwarding, and authenticated SOCKS5 proxy support while continuing to perform phishing overlays, keylogging, screen recording, SMS interception, one time password suppression, clipboard modification, notification filtering, and screenshot capture.
8 guiding principles for reskilling the SOC for agentic AI

www.csoonline.com · 5 months ago

There are eight guiding principles for reskilling security operations for agentic AI, including embracing the agentic imperative, setting the tone from the top, respecting resistance while overcoming it, and providing hands-on, sandboxed training. Security leaders report embedding vendor experts, running bootcamps and internal sandboxes such as LabX to encourage rapid experimentation while redeploying humans into higher-value oversight and automation design roles. Organizations are urged to emphasize governance and human-in-the-loop controls with escalation paths, audit trails and override authorities, to rethink operating models and career ladders as L1 and L2 functions become agentic, and to train analysts in model intuition and policy setting so agents operate within defined decision boundaries.
See the hidden rules behind AI. Then use them to rewrite this article.

www.washingtonpost.com · 5 months ago

There are thousands of words of hidden instructions called system prompts that companies inject into every chatbot conversation to steer personality, policy compliance, and tool use, and those instructions can take priority over what a user types. Researchers who extracted system prompts found they range from a few hundred to more than twenty thousand words and include specific rules such as forbidding unnecessary discussion of goblins and strict limits on quoting copyrighted material. Companies generally keep full system prompts secret and offer only limited personalization features that let users request tone or formatting changes but not directly edit the underlying instructions.
1,800+ MCP servers exposed without authentication: How zero trust can secure the AI agent revolution

www.csoonline.com · 5 months ago

There are widespread exposures of Model Context Protocol servers, with 1,862 publicly reachable instances identified and a verified sample of 119 that all permitted unauthenticated access to internal tool listings. Exposed servers include production systems with write access to financial databases, social media accounts, and customer relationship management platforms, creating opportunities for tool poisoning, rug pulls, and cross-server contamination that enable stealthy data exfiltration. Researchers have documented high-impact vulnerabilities, including EchoLeak CVE-2025-32711, a zero-click technique that hides malicious instructions in documents to coerce Microsoft 365 Copilot into exfiltrating data, and mcp-remote CVE-2025-6514, a JFrog package command injection via unsanitized OAuth parameters that can yield full system compromise. Security recommendations include adopting the Agentic Trust Framework and zero-trust controls such as identity governance for agents, mandatory authentication and authorization for every interaction, cryptographic signing and immutable versioning of tool definitions, network segmentation, continuous behavioral monitoring, and human-in-the-loop approval for sensitive operations.
Canvas System Is Online After a Cyberattack Disrupted Thousands of Schools

www.securityweek.com · 5 months ago

On May 11, 2026, tens of thousands of students regained access to the Canvas online learning system after a cyberattack had knocked the platform offline during finals, forcing schools to reschedule exams and seek workarounds. Instructure reported that an unauthorized actor made changes to pages visible to logged in users and that the attacker exploited an issue related to Free For Teacher accounts, which the company temporarily disabled while investigating. A hacking collective claiming to be ShinyHunters said nearly nine thousand schools were affected and that billions of private messages and records were accessed, and Instructure has not confirmed whether a ransom was paid or which data was compromised.
OpenClaw Malware Targets Crypto Wallets and Bitwarden Credentials

gbhackers.com · 5 months ago

OpenClaw users are being targeted by a malware campaign that distributes a fake OpenClaw Windows installer containing a 130 megabyte Rust executable named Hologram padded with fake documentation to evade antivirus and sandbox file size limits. After multi layer anti virtual machine checks that can wait for real mouse movement, the dropper decodes and runs an obfuscated PowerShell payload that disables Microsoft Defender, opens inbound firewall ports, retrieves a password protected archive and drops multiple Rust based modules that establish resilient persistence and Telegram based callbacks via a Hookdeck relay. The stage two Stealth Packer framework includes six Rust components that execute payloads entirely in memory using techniques such as hosting the.NET CLR inside Rust and reflectively executing decrypted PE files via direct NT syscalls, and it uses a server side manifest to target 250 browser extensions including 201 crypto wallet extensions such as MetaMask, Phantom, Coinbase and OKX, plus 49 password managers including Bitwarden, LastPass and 1Password. Operators also deploy a separate Ledger Live module to steal funds and rely on legitimate cloud services including Azure DevOps, Telegram and Hookdeck for command and control, making detection difficult and requiring defenders to inspect application layer traffic rather than rely on IP only perimeter controls.
MyAudi app:Security issues in Audi Connected Vehicle experience

decoder.cloud · 5 months ago

A security researcher found that the myAudi connected vehicle platform allows anyone who knows a vehicle identification number to add that vehicle as a guest user and access sensitive vehicle and ownership information. Some GraphQL endpoints had introspection enabled exposing API schemas, and vehicle management endpoints returned embedded SIM identifiers such as IMEI and ICCID along with license and service metadata. Guest users could poll and retrieve pending requests initiated by primary users, including honk and flash responses that contained the requesting user's GPS location and unlock status; that specific issue has been fixed following disclosure while other findings remain under evaluation and a full review of vehicle association logic was recommended.
New ‘Dirty Frag’ Linux Vulnerability Possibly Exploited in Attacks

www.securityweek.com · 5 months ago

A newly disclosed local privilege escalation vulnerability affecting major Linux distributions, called Dirty Frag and Copy Fail 2 and tracked as CVE-2026-43284 and CVE-2026-43500, can allow an unprivileged user to escalate permissions to root. The exploit chains two kernel flaws affecting the xfrm-ESP and RxRPC components, was publicly disclosed before patches were available, and technical details and proof of concept code were made available online. Microsoft and other vendors report limited in the wild activity that may indicate exploitation, including post compromise modification of authentication and PHP session files. Linux distributors including Red Hat, Amazon Linux, Ubuntu, Fedora and AlmaLinux have released patches and mitigations and organizations are advised to apply updates.
Giving Claude Code Full Control of a Hardware Fault Injection Setup to Bypass Secure Boot

raelize.com · 5 months ago

Researchers used an autonomous Claude Code agent to reproduce a fault injection attack that bypassed Secure Boot V1 on an Espressif ESP32 system on chip using voltage crowbar glitches. The AI wrote all of the software tooling and orchestrated hardware including a ChipWhisperer Husky, an RK6006 power supply, a PicoScope, and an ESP-PROG adapter, reverse-engineered and emulated the rev-0 boot ROM, ran thousands of glitch attempts, and built a live monitoring dashboard while the campaign ran. The issue is specific to ESP32 Secure Boot V1 and was mitigated in ESP32 V3 by modifying ROM code, though related fault injection risks remain.
Crimenetwork Bust Reveals 22,000 Members and Over 100 Illicit Vendors

gbhackers.com · 5 months ago

Law enforcement dismantled the relaunched Crimenetwork online marketplace after arresting a 35 year old German national at his residence in Mallorca, Spain on a European Arrest Warrant. Investigators seized about 194,000 euros in physical and digital assets, recovered user databases and transaction logs, and found the site had grown to more than 22,000 registered users and over 100 active vendors trading stolen data, narcotics and forged identification. The marketplace used Bitcoin, Litecoin and Monero to process payments and generated more than 3.6 million euros in revenue while charging commissions and recurring vendor fees. The operation was a coordinated effort between German, Spanish, Moldovan and European agencies and follows a March 2026 conviction that resulted in a seven year ten month prison sentence and the confiscation of over ten million euros for the previous operator.