Blahalla

PyPI Telnyx Python SDK Backdoored to Steal Credentials on Windows, macOS, and Linux

gbhackers.com · 3 months ago

A backdoored Telnyx Python SDK published to PyPI used a multi stage WAV steganography payload to steal credentials on Windows, macOS, and Linux, with malicious code placed in the package client module so that simply importing Telnyx executes the payload before application code runs. Two malicious releases, versions 4.87.1 and 4.87.2, were published on March 27, 2026 without matching GitHub commits and remained available for about six and a half hours before PyPI quarantined them, and users are advised to revert to version 4.87.0 and assume systems that installed those versions are fully compromised. The attackers split obfuscated Base64 blobs across the module and fetch seemingly valid WAV files at runtime, decode and XOR decrypt eight byte keyed payloads to reconstruct credential stealers that on Windows drop a msbuild executable into user startup with a hidden lock to prevent replay, while on Linux and macOS the stealer runs detached and exfiltrates SSH keys, environment variables, cloud credentials and other secrets in encrypted archives. Attribution to the TeamPCP actor is supported by reuse of an RSA 4096 key, a tpcp dot tar dot gz campaign identifier and a custom exfiltration header, and defenders are urged to hunt for unusual WAV downloads and outbound HTTP posts carrying tpcp dot tar dot gz, remove suspicious msbuild executables and hidden lock files, rotate all exposed credentials, and pin, hash and restrict PyPI publishing to prevent future supply chain compromise.

Google Drive ransomware detection now on by default for paying users

www.bleepingcomputer.com · 3 months ago

Google announced that its AI powered Google Drive ransomware detection feature has reached general availability and is now enabled by default for paying customers. When ransomware encrypted files are detected during a desktop sync, Drive immediately pauses syncing, notifies the affected user and administrators, and creates an alert in the Admin console. While files on the compromised endpoint may still be encrypted, documents stored in Drive are protected and can be restored using the Drive restoration tool, and Google says its latest AI model detects 14 times more infections than during the beta. Administrators can disable the feature in the Admin console and detection alerts require installing Drive for desktop version 114 or later on endpoints, although syncing will still pause on older versions.

TrueConf Vulnerability Under Active Exploitation in Southeast Asia Government Attacks

gbhackers.com · 3 months ago

Check Point Research discovered a critical zero day vulnerability in the TrueConf video conferencing client. The flaw is tracked as CVE-2026-3502 with a CVSS score of 7.8 and is being actively exploited in targeted attacks against government entities in Southeast Asia. The campaign called Operation TrueChaos leverages the TrueConf trusted update mechanism to deliver the Havoc post exploitation framework.

Claude Code Source Leaked via npm Packaging Error, Anthropic Confirms

thehackernews.com · 3 months ago

Internal source code for the Claude Code AI assistant was inadvertently published when an npm release, version 2.1.88, included a source map that exposed nearly two thousand TypeScript files and more than 512,000 lines of code. Anthropic said no sensitive customer data or credentials were exposed and attributed the incident to a release packaging error caused by human error. Users who installed or updated via npm on March 31, 2026 between 00:21 and 03:29 UTC may also have pulled a trojanized HTTP client tied to a separate supply chain incident, and security researchers warn that typosquat packages were published as empty stubs that could be weaponized later. The leak revealed internal features including a persistent background agent called KAIROS, a dream mode, self healing memory architecture, multi agent orchestration, and an Undercover Mode, creating a blueprint attackers can use to craft targeted jailbreaks, dependency confusion, or supply chain attacks.

NPM Supply Chain Attack Uses undicy-http to Deploy RAT

gbhackers.com · 3 months ago

A sophisticated supply chain attack targeted Node.js projects through a malicious npm package named undicy-http, impersonating the legitimate undici HTTP client. The package delivers a remote access trojan, or RAT, and a browser stealer, with ties to the LofyGang threat group. The malware employs techniques to evade detection, including hidden execution via VBScript, process hollowing using native syscalls, and anti-virtual machine checks. It steals credentials, cookies, and data from over 50 browsers and numerous wallet extensions, exfiltrating data through Discord and Telegram. Systems potentially impacted should undergo complete remediation to mitigate the compromise.

New Windows 11 emergency update fixes preview update install issues

www.bleepingcomputer.com · 3 months ago

Microsoft has released an emergency update, KB5086672, for Windows 11 versions 24H2 and 25H2 to resolve installation problems with a previous preview update, KB5079391. Users were experiencing error code 0x80073712 and missing files during the update process, prompting Microsoft to halt the rollout. This new update supersedes the prior one and includes all previous protections and improvements. Earlier this month, Microsoft also released updates to fix Microsoft account sign-in issues and address vulnerabilities in the Routing and Remote Access Service.

Google Cloud’s Vertex AI Hit by Vulnerability Enabling Sensitive Data Access

gbhackers.com · 3 months ago

Security researchers have discovered a vulnerability in Google Cloud’s Vertex AI Agent Engine that allows attackers to access sensitive data and compromise infrastructure. The vulnerability stems from overly broad default permissions, allowing the deployment of a malicious AI agent that can steal service agent credentials. Attackers can then use these credentials to access Google Cloud Storage Buckets, Artifact Registry repositories, and deployment files, potentially impacting Google Workspace data. Google has updated its documentation and recommends organizations adopt a Bring Your Own Service Account architecture to enforce strict access controls.

Hackers Actively Exploit Critical WebLogic RCE Vulnerabilities in Ongoing Attacks

gbhackers.com · 3 months ago

Attackers are actively exploiting a critical unauthenticated Remote Code Execution vulnerability, CVE-2026-21962, in Oracle WebLogic Server, with exploitation beginning immediately after public exploit code was released. Researchers observed automated scanning and attacks originating from rented Virtual Private Servers, utilizing tools like libredtail-http and Nmap. Besides the new vulnerability, attackers are also persistently targeting older WebLogic flaws, including CVE-2020-14882, CVE-2020-2551, and CVE-2017-10271, highlighting the need for urgent patching and security measures. Organizations are advised to restrict console access, implement firewalls, and enhance log monitoring to mitigate these risks.

XLoader malware Sharpens Obfuscation, Masks C2 Traffic via Decoy Servers

gbhackers.com · 3 months ago

XLoader malware has evolved with new versions, particularly 8.1 and 8.7, to significantly harden its code and obscure command-and-control traffic. Developers have layered obfuscation techniques, including encrypted strings, custom RC4-based decryption routines, and randomized HTTP parameters, to evade detection. The malware now uses multiple encryption keys and decoy infrastructure, embedding 65 encrypted C2 IP addresses that are decrypted at runtime. XLoader’s capabilities include credential theft, remote command execution, and the ability to execute PowerShell scripts and PE files.

Market Bifurcation in Pentesting by 2026 (37%) – AI May Split the Field in Two Rather Than Flatten It, and That Changes Everything About Who Survives the Disruption

usemurmur.ai · 3 months ago

false

Anthropic employee error exposes Claude Code source

www.csoonline.com · 3 months ago

An Anthropic employee inadvertently exposed the source code for the AI coding tool, Claude Code, by including a source map file in a publicly accessible npm registry. Source map files act as a bridge between minified code and the original source code, allowing developers to debug but potentially revealing proprietary information and vulnerabilities if exposed. Anthropic stated the incident was due to human error and involved no customer data, and they are implementing preventative measures. Experts highlight this as a common mistake with serious implications due to the high value of the intellectual property and potential for attackers to identify vulnerabilities.

Claude Code source code accidentally leaked in NPM package

www.bleepingcomputer.com · 3 months ago

Anthropic accidentally leaked the closed-source code for Claude Code through a packaging error when releasing version 2.1.88 on NPM. The leak included a 60 megabyte source map file containing approximately 500,000 lines of code and revealing details about Claude-exclusive features, such as a "Proactive mode" and a "Dream" mode. Anthropic is attempting to remove the leaked code through DMCA infringement notices and is investigating a bug causing Claude Code usage limits to deplete faster than anticipated. The company confirms no customer data or credentials were exposed during the incident.

Google now allows you to change your @gmail.com address

www.bleepingcomputer.com · 3 months ago

false

Proton launches new "Meet" privacy-focused conferencing platform

www.bleepingcomputer.com · 3 months ago

false

GIGABYTE Control Center vulnerable to arbitrary file write flaw

www.bleepingcomputer.com · 3 months ago

A security vulnerability has been identified in GIGABYTE Control Center, a utility pre-installed on GIGABYTE laptops and motherboards. The flaw allows unauthenticated attackers to write files to a system, potentially leading to code execution and privilege escalation. Versions 25.07.21.01 and earlier are affected if the “pairing” feature is enabled. The vulnerability, tracked as CVE-2026-4415, has a critical severity rating and users are urged to update to version 25.12.10.01.

The Forgotten Endpoint: Security Risks of Dormant Devices

www.darkreading.com · 3 months ago

There are widespread risks from forgotten or dormant corporate devices, such as contractor laptops that retain VPN access, saved credentials, and certificates and thus provide direct paths into enterprise networks. The problem stems from poor asset inventory and endpoint visibility, which undermines zero trust, creates compliance failures under frameworks such as HIPAA and NIST SP 800-53 CM-8, and produces financial and security impacts; a Kensington study found 76 percent of IT decision makers reported device theft in the past two years and 46 percent experienced a breach due to stolen or unsecured devices. Recommended mitigations include shifting contractors to bring your own device with access via virtual desktops or cloud workspaces, automating discovery and flagging devices dormant for more than 45 days, enforcing remote wipe and incident response playbooks for stolen or rogue devices, and conducting regular audits to restore visibility.

Claude AI finds Vim, Emacs RCE bugs that trigger on file open

www.bleepingcomputer.com · 3 months ago

Researchers have discovered remote code execution vulnerabilities in the Vim and GNU Emacs text editors. The vulnerabilities in Vim allow malicious code to be executed simply by opening a specially crafted file, and a patch has been released. The GNU Emacs vulnerability arises from its integration with Git, allowing arbitrary command execution when opening a file containing a malicious.git/config file. This flaw remains unpatched, and users are advised to exercise caution when opening files from untrusted sources.

OpenAI Adds Another $12 Billion to Latest Funding Round

www.nytimes.com · 3 months ago

false

New North Korean AI Hiring Scheme Targets US Companies

www.techrepublic.com · 3 months ago

North Korean operatives are using sophisticated techniques to infiltrate companies through their hiring processes. These schemes combine stolen identities, AI-generated resumes tailored to job descriptions, and anonymized infrastructure like VPNs and laptop farms to bypass standard screening. During interviews, candidates may exhibit unusual behavior suggesting reliance on AI chatbots and struggle with unexpected questions. Organizations are advised to implement layered security measures, including identity verification, live work demonstrations, and device controls, to mitigate this growing threat.

Weaponizing the Protectors: TeamPCP’s Multi-Stage Supply Chain Attack on Security Infrastructure

unit42.paloaltonetworks.com · 3 months ago

A threat group, TeamPCP, has launched a supply chain attack targeting several open-source security tools and Python packages. They compromised publishing credentials to inject malicious code and steal sensitive data such as SSH keys and cloud credentials. The group is now partnering with ransomware groups and using techniques like CanisterWorm for self-replication and data wiping, potentially impacting 500,000 machines. Palo Alto Networks recommends auditing CI/CD pipelines and prioritizing SBOM visibility for mitigation.

Axios NPM Package Compromised in Precision Attack

www.darkreading.com · 3 months ago

The Axios JavaScript package, a widely used HTTP client library, was recently compromised in a sophisticated supply chain attack. Malicious versions of the package contained a hidden, remote-access Trojan disguised as a legitimate crypto-js library. The attack, which lasted approximately three hours, involved staging a malicious dependency for 18 hours prior to release and involved the compromise of the lead maintainer's account. Google attributes the attack to the suspected North Korean threat actor UNC1069, suggesting motives of access brokering or espionage.

Attackers trojanize Axios HTTP library in highest-impact npm supply chain attack

www.csoonline.com · 3 months ago

Attackers compromised the npm account of the lead maintainer of Axios, a widely used JavaScript HTTP client library, and published malicious versions, [email protected] and [email protected], that deployed a cross-platform remote access trojan. The incident, attributed to a North Korean threat actor, UNC1069, affected approximately 3% of impacted environments and rippled through the npm ecosystem, impacting almost 175,000 projects. The malware, delivered through a postinstall hook in a decoy package called [email protected], downloaded platform-specific RAT payloads and attempted to erase its tracks. Security firms quickly detected and removed the malicious packages, but organizations should audit dependencies, rotate credentials, and consider implementing stricter npm policies to mitigate the risk of similar attacks.

Google's Vertex AI Is Over-Privileged. That's a Problem

www.darkreading.com · 3 months ago

Researchers at Palo Alto Networks have identified a significant security risk within Google's Vertex AI platform. The platform's default service account, the Per-Project, Per-Product Service Agent, grants excessive permissions that attackers can exploit to steal data, access internal infrastructure, and potentially execute unauthorized actions. Google has updated documentation and recommends organizations replace the default service agent with a custom service account to enforce least-privilege access. The risk highlights the need for organizations to prioritize AI security and control agent permissions to prevent unauthorized actions.

TeamPCP Breaches Cloud, SaaS Instances With Stolen Credentials

www.darkreading.com · 3 months ago

The threat group TeamPCP is exploiting supply chain attacks against open source software projects to steal credentials and gain access to cloud and SaaS environments. They initially compromised projects including Trivy, KICS, LiteLLM, and Telnyx to deploy malware and harvest sensitive information. Following the credential theft, TeamPCP rapidly breached AWS and Azure environments, exfiltrating data and executing commands within compromised containers. Organizations are urged to rotate credentials and monitor for suspicious activity such as unusual VPN use and excessive git clone events.

MAD Bugs: Claude Wrote a Full FreeBSD Remote Kernel RCE with Root Shell (CVE-2026-4747)

blog.calif.io · 3 months ago

Researchers used the AI model Claude to develop a remote code execution exploit for FreeBSD, resulting in a root shell. The process, which took approximately eight hours of wall clock time, involved solving six distinct challenges. The exploit utilized multiple packets to write shellcode to kernel memory and leveraged techniques to manage kernel threads and transition to user mode. This development, attributed to Nicholas Carlini using Claude, is identified as CVE-2026-4747 and demonstrates a capability previously thought to be exclusive to human expertise.

PSA: That 'Disable NTLMv1' GPO you set years ago? It’s lying to you. LmCompatibilityLevel set to 5 is not enough.

www.silverfort.com · 3 months ago

A critical security issue has been identified in Active Directory environments where organizations mistakenly believe NTLMv1 authentication is disabled. Misconfigured on-premise applications can bypass these security measures, allowing attackers to steal credentials and elevate privileges, potentially impacting a large percentage of accounts. Experts recommend auditing NTLM usage, implementing security fencing, and migrating to more secure authentication methods, as Microsoft plans to remove NTLMv1 entirely. Organizations are encouraged to proactively detect and mitigate this ongoing threat through application mapping and modern security implementations.

Cisco Nexus Dashboard Fabric Controller Arbitrary Command Execution Vulnerability

sec.cloudapps.cisco.com · 3 months ago

A vulnerability has been identified in the Cisco Nexus Dashboard Fabric Controller, allowing an authenticated attacker to execute commands with network-admin privileges. The issue stems from improper user authorization and insufficient command argument validation. Cisco recommends upgrading to a fixed software release to remediate this vulnerability, as no workarounds are available. The vulnerability does not affect configurations used for storage area network deployments.

5-month-old F5 BIG-IP DoS bug becomes critical RCE exploited in the wild

www.csoonline.com · 3 months ago

A critical remote code execution vulnerability, CVE-2025-53521, has been discovered in F5 BIG-IP Access Policy Manager (APM) and is currently being exploited. Initially classified as a denial-of-service issue, the flaw allows attackers to gain root-level access and deploy persistent malware. Affected versions include BIG-IP APM 17.1.0 to 17.1.2, 17.5.0 to 17.5.1, 16.1.0 to 16.1.6, and 15.1.0 to 15.1.10, with patches available in newer versions. Organizations are urged to immediately patch vulnerable systems and assess for potential compromise, even if prior updates were applied.

Android Developer Verification Rollout Begins Ahead of September Enforcement

thehackernews.com · 3 months ago

Google is rolling out Android developer verification to combat malicious apps distributed anonymously. This verification requires developers distributing apps outside of the Google Play Store to confirm their identity through the Android Developer Console. Users will experience no change when installing apps, but power users sideloading unregistered APKs will encounter an authentication step and a 24-hour waiting period. The move follows Apple’s recent revisions to its Developer Program License Agreement, restricting third-party wearables' access to user data.

Google Drive Expands AI Ransomware Detection, File Recovery to More Users

www.techrepublic.com · 3 months ago

Google is expanding ransomware detection and file recovery capabilities within Google Drive. A new artificial intelligence model now identifies 14 times more infections than previously available. When ransomware is detected on a desktop device, file syncing pauses to prevent further damage and alerts are sent to users and administrators. File restoration is now available to all Google Workspace customers, Workspace Individual subscribers, and personal Google accounts, while ransomware detection is available to select Google Workspace tiers.

Cisco source code stolen in Trivy-linked dev environment breach

www.bleepingcomputer.com · 3 months ago

Cisco recently experienced a cyberattack where threat actors exploited stolen credentials from the Trivy supply chain compromise to breach its internal development environment. The attackers used a malicious GitHub Action plugin to steal credentials and data, impacting numerous devices and compromising over 300 GitHub repositories containing source code for Cisco products and customer data. Multiple AWS keys were also stolen and used for unauthorized activity, and the company is now re-imaging systems and rotating credentials. Security researchers have linked these attacks to the TeamPCP group, who have also targeted LiteLLM and Checkmarx.

Democrats Examine Elon Musk’s Role in Suspension of Business Disclosure Law

www.nytimes.com · 3 months ago

false

Lesser-Known Military College Triumphs in Pentagon Student Hacking Contest

www.govinfosecurity.com · 3 months ago

The University of North Georgia’s cybersecurity team won the inaugural Cyber Workforce Competition, defeating teams from military colleges like The Citadel. The competition, modeled after March Madness, involved teams racing to complete objectives in a simulated environment, with the NSA providing feedback and monitoring stealth. The challenges included geolocating a high value individual, spreading malware through a network, and penetrating connected systems while avoiding detection. Organizers plan to expand the competition next year to include more universities.

The threat to critical infrastructure has changed. Has your readiness?

www.microsoft.com · 3 months ago

Critical infrastructure organizations face a significantly evolving cyber threat landscape, with attackers establishing persistent footholds for future disruption. Governments worldwide are implementing stricter cybersecurity regulations, including the U.S. National Cybersecurity Strategy, Japan’s Active Cyber Defense legislation, Europe’s NIS2 Directive, and Canada’s Bill C8. Practical training and coaching are vital to improve cyber readiness, particularly in sectors like water and wastewater utilities. Investigations reveal a trend of identity-driven intrusion, living-off-the-land persistence, and nation-state prepositioning within hybrid IT-OT environments.

ImageMagick: From Arbitrary File Read to File Write In Every Policy (ZeroDay)

pwn.ai · 3 months ago

Security researchers have identified critical vulnerabilities in ImageMagick, a widely used image processing library, impacting systems like WordPress and various Linux distributions. These zero-day exploits, including CVE-2024-24421, allow for file reads, writes, and potential remote code execution by bypassing security policies and exploiting weaknesses in format detection and delegate handling. The vulnerabilities stem from a combination of ImageMagick, Ghostscript, and default server configurations, and mitigation strategies include disabling Ghostscript and upgrading ImageMagick. Concerns about the security reporting process led to the public release of these findings.

TrueConf Zero-Day Exploited in Attacks on Southeast Asian Government Networks

thehackernews.com · 3 months ago

Cybersecurity researchers have identified a zero-day exploit targeting government networks in Southeast Asia, in a campaign called TrueChaos. The vulnerability, CVE-2026-3502, allows attackers to distribute malicious software updates by compromising the TrueConf video conferencing server. This enables the deployment of the Havoc command-and-control framework and DLL backdoors, with links to a Chinese-nexus threat actor. The vulnerability has been patched in version 8.5.3 of the TrueConf Windows client.

Applying security fundamentals to AI: Practical advice for CISOs

www.microsoft.com · 3 months ago

false

Censys Raises $70 Million for Internet Intelligence Platform

www.securityweek.com · 3 months ago

false

Rethinking Vulnerability Management Strategies for Mid-Market Security

www.darkreading.com · 3 months ago

Security teams are increasingly challenged by a rapidly growing number of vulnerabilities, now reaching 50,000 annually. Simply counting vulnerabilities isn’t enough; the speed of remediation is critical, as the mean time to exploit has decreased to just hours. Many organizations remain vulnerable despite patching due to overlooked misconfigurations and broader attack surfaces. Experts recommend focusing on attack surface management rather than solely relying on CVE data to address these risks.

The Next Cybersecurity Crisis Isn’t Breaches—It’s Data You Can’t Trust

www.securityweek.com · 3 months ago

false

AI and Quantum Are Forcing a Rethink of Digital Trust

www.darkreading.com · 3 months ago

false

Stolen Logins Are Fueling Everything From Ransomware to Nation-State Cyberattacks

www.securityweek.com · 3 months ago

Stolen credentials are a primary driver of escalating cyberattacks, ranging from ransomware to nation-state operations, with credential resale occurring on a large scale. Attackers are increasingly employing double or triple extortion tactics, including data theft and operational disruption, alongside the use of artificial intelligence to develop malware. These tactics are fueling supply chain attacks and expanding to include civilian entities, often with geopolitical motivations, as seen in incidents such as the Lazarus Group cryptocurrency theft and DDoS attacks exceeding 31.4 Tbps. Organizations must shift focus from preventing credential theft to recognizing and blocking misuse through adaptive identity measures.

Venom Stealer Raises Stakes With Continuous Credential Harvesting

www.securityweek.com · 3 months ago

A new malware-as-a-service kit called Venom Stealer has emerged, offering credential harvesting capabilities through a licensing model. The kit, operated via Telegram, uses social engineering lures like fake Cloudflare CAPTCHAs to install a Windows payload that targets Chromium and Firefox browsers, extracting sensitive data like passwords and cryptocurrency wallet vaults. Uniquely, Venom Stealer establishes persistent monitoring, continuously capturing new credentials and wallet activity, and automatically cracks several cryptocurrency wallets, including Tonkeeper. The kit's developer added the Chrome v10/v20 bypass and auto-crack support in March 2026.

I broke up with my Kindle. My new e-reader treats me better.

www.washingtonpost.com · 3 months ago

false

Common Entra ID Security Assessment Findings – Part 2: Privileged Unprotected Groups

blog.compass-security.com · 3 months ago

Microsoft Entra ID faces a security risk from improperly configured groups, allowing unauthorized users and applications to modify group membership and escalate privileges. An Azure Arc vulnerability could allow attackers to execute commands as the SYSTEM account on on-premises Active Directory servers. To address the Entra ID risk, the open-source PowerShell tool EntraFalcon helps identify unprotected groups. Securing group configurations, using Restricted Management Administrative Units, or making groups role-assignable are recommended remediation steps.

How to Categorize AI Agents and Prioritize Risk

www.bleepingcomputer.com · 3 months ago

false

Hackers compromise Axios npm package to drop cross-platform malware

www.bleepingcomputer.com · 3 months ago

Hackers compromised the npm account for the Axios JavaScript package, a widely used HTTP client, to distribute malware across Linux, Windows, and macOS systems. The attackers injected a malicious dependency that downloads and executes platform-specific payloads, installing a remote access trojan (RAT). This attack involved staged preparation and self-deleting actions to evade detection, but is not attributed to the TeamPCP group. Users are advised to lock their Axios versions to 1.14.0 and 0.30.3 and rebuild environments if a compromise is suspected.

TeamPCP Moves From OSS to AWS Environments

www.securityweek.com · 3 months ago

A threat actor group, known as TeamPCP, has expanded its operations following a recent breach of the Aqua Security Trivy vulnerability scanner. Initially focused on cloud environments and supply chain attacks, the group is now compromising AWS environments by using stolen credentials to access and exfiltrate data. Researchers report that the group validates stolen secrets like AWS access keys and then performs discovery operations, targeting containers and AWS Secrets Manager. They have also reportedly shared compromised data and secrets with groups like Lapsus$ and the Vect Ransomware Group.

WhatsApp malware campaign delivers VBScript and MSI backdoors

www.microsoft.com · 3 months ago

A recent malware campaign used WhatsApp messages to distribute malicious Visual Basic Script files, leading to the installation of unsigned MSI packages. Attackers disguised legitimate Windows tools and used cloud services such as Amazon S3 and Backblaze B2 to avoid detection. The attack involved multiple steps, including initial access, privilege escalation, and payload delivery. Microsoft recommends strengthening endpoint controls and user education to defend against this threat.

CrewAI Vulnerabilities Expose Devices to Hacking

www.securityweek.com · 3 months ago

Security researchers have identified four vulnerabilities in the CrewAI multi-agent orchestration framework that could allow attackers to execute code, steal credentials, and access sensitive files. The issues, including CVE-2026-2275, CVE-2026-2286, CVE-2026-2287, and CVE-2026-2285, are linked to the Code Interpreter tool and improper default configurations. Attackers can chain these vulnerabilities through prompt injections to escape the intended sandbox environment. Developers are advised to restrict the Code Interpreter tool, implement secure configurations, and sanitize user input to mitigate the risks.