CVE-2026-2796 Prompt Injection Vulnerability affecting Claude

research.aimultiple.com · Jan 29, 2026 05:30

The article highlights advanced threats to AI agents, including "Shadow Escape," a zero-click exploit targeting Model Context Protocol (MCP) based systems that enables workflow hijacking and data exfiltration. Another critical vulnerability involves "ASCII Smuggling," where hidden malicious prompts embedded in files lead to remote code execution and sensitive data exfiltration by deceiving LLMs like Claude.

thehackernews.com · Feb 06, 2026 05:30

Anthropic's Claude Opus 4.6 LLM has identified over 500 previously unknown, high-severity security vulnerabilities, including memory corruption and buffer overflow issues, in critical open-source libraries like Ghostscript, OpenSC, and CGIF. This demonstrates AI's emerging capability for sophisticated vulnerability discovery and code analysis, even for complex flaws requiring conceptual understanding of algorithms.

venturebeat.com · Feb 10, 2026 05:30

Anthropic's Claude Opus 4.6 exhibits prompt injection success rates up to 78.6% in less constrained environments, quantitatively validating a previously theoretical risk for AI agents. This vulnerability was demonstrated by a PromptArmor attack on Claude Cowork, enabling data exfiltration of confidential files via hidden prompt injections bypassing sandbox restrictions and monitoring.

anthropic.com · Mar 06, 2026 05:30

Anthropic's Claude Opus 4.6 AI model discovered 22 novel vulnerabilities in Firefox, 14 of which were high-severity, leading to fixes in Firefox 148.0 for hundreds of millions of users. The AI also demonstrated the ability to automatically develop crude browser exploits for some of these vulnerabilities, underscoring the potential for AI in accelerated vulnerability discovery and exploit generation.

thehackernews.com · Mar 07, 2026 05:30

Anthropic's Claude Opus 4.6 AI model discovered 22 new vulnerabilities in the Firefox browser, including high-severity issues like a use-after-free bug and a critical just-in-time miscompilation. The AI also demonstrated the concerning ability to develop crude exploits for identified flaws, specifically one for CVE-2026-2796 with a CVSS score of 9.8.

sentinelone.com · Mar 31, 2026 05:30

A sophisticated multi-stage supply chain attack, initiated by compromising open-source security scanner Trivy to steal LiteLLM PyPI credentials, injected malicious versions (1.82.7, 1.82.8) of LiteLLM into customer environments. This enabled data exfiltration, system persistence, and lateral movement within Kubernetes clusters, notably leveraging AI coding assistants with unrestricted permissions as an unwitting infection vector.

fortune.com · Mar 31, 2026 05:30

Anthropic accidentally leaked the source code for its Claude Code AI coding tool's agentic harness, comprising 500,000 lines of code across 1,900 files, via an improper NPM package release. This exposure allows for reverse engineering of the AI's core logic, revealing internal APIs and potentially enabling sophisticated actors to bypass model safeguards or understand its architecture.

theguardian.com · Apr 01, 2026 05:30

Anthropic accidentally exposed nearly 2,000 internal files and 500,000 lines of source code for its AI-powered coding assistant, Claude Code, due to human error during a software update. This disclosure revealed internal architecture blueprints and commercially sensitive operational instructions for their AI models, raising competitive and internal security concerns.

thehackernews.com · Apr 01, 2026 05:30

The Claude Code source code was inadvertently leaked due to an npm packaging error, specifically exposing thousands of TypeScript files via a source map file in version 2.1.88. This leak directly enabled bad actors to bypass AI guardrails, conduct supply chain attacks via trojanized npm packages like Axios, and distribute malware such as Vidar Stealer through dependency confusion and fake GitHub repositories.

fortune.com · Apr 02, 2026 05:30

A supply-chain cyberattack on the open-source LiteLLM library led to the planting of malicious code designed for credential harvesting. This incident resulted in a significant data breach at AI startup Mercor, potentially exposing sensitive company data, user information, and confidential AI project details.

techbuzz.ai · Apr 04, 2026 05:30

Threat actors are weaponizing leaked Anthropic Claude AI source code by embedding malware, disguised as legitimate repositories, and distributing it to developers. This malicious distribution aims to install backdoors and credential harvesters on victims' systems, providing attackers with potential footholds into sensitive corporate and research networks.

businessinsider.com · Apr 04, 2026 05:30

AI training startup Mercor experienced a data breach resulting from a supply chain attack that leveraged the open-source project LiteLLM, impacting potentially thousands of companies. This incident prompted Meta to pause its collaboration with Mercor while a thorough investigation, supported by third-party forensics experts, is conducted.

helpnetsecurity.com · Aug 13, 2026 05:30

A supply chain attack compromised LiteLLM via a poisoned Trivy dependency, leading to the deployment of malicious LiteLLM versions (1.82.7/1.82.8) that exfiltrated 153GB of corporate credentials. This dataset, comprising AWS keys, API tokens, and other secrets from CI runner environments, impacts nearly 2,500 organizations, underscoring critical software supply chain vulnerabilities.