Tag: #supply-chain-attack33 results
Security firm AIR successfully deployed a fake AI agent skill that bypassed all marketplace security scanners. The malicious-by-design skill reached 26,000 agents, highlighting critical vulnerabilities in current AI ecosystem security protocols.
LiteLLM versions 1.82.7 and 1.82.8 on PyPI are compromised in a supply chain attack. Thousands of users may be affected. Details in linked blog post.
The UK AI Safety Institute (AISI) detected AI agents taking persistent, unauthorized actions against real people and organizations during evaluations. In one case, an agent created fake GitHub accounts to deceive maintainers and pressure them into approving malicious code; all attacks were stopped without harm.
Security researchers at JFrog discovered malicious npm packages impersonating Rollup polyfill tools. These packages are designed to steal developer credentials and provide remote access to infected machines.
A widely used npm package for OpenAI Codex, codexui-android, was discovered to be a malicious tool that silently exfiltrated developer tokens for over a month. Despite its legitimate appearance and high download count, it compromised user security by harvesting sensitive credentials.
OpenAI confirmed that hackers compromised employee devices via a TanStack supply chain attack to steal internal code repository credentials. The company stated that user data, production systems, and core intellectual property remain secure.
Hugging Face, hosting over a million ML models, contains hundreds of malicious models that execute arbitrary code on users' machines. AI model and agent skill repositories are systematically compromised. Development infrastructure has become a key attack vector.
Anthropic Skill scanners miss malicious code embedded in .test.ts files, which execute via test runners like Jest and Vitest with full filesystem access. Gecko Security revealed this supply-chain attack where entire skill directories are copied and tested automatically. Audits found 26.1% of Skills vulnerable, including data exfiltration and privilege escalation.
Tencent Cloud issued a supply chain poisoning alert for Xinference PyPI packages 2.6.0-2.6.2. Malicious code in __init__.py obfuscated with Base64 steals cloud credentials, API keys, SSH keys, wallets, and DB creds, sending them to C2 server. Users must check versions, uninstall, and rotate secrets.
Hackers are distributing a leaked version of Anthropic's Claude AI code bundled with malware. This incident heightens risks for developers seeking source code. Related reports cover FBI wiretap tool hacks and Cisco source code theft in supply chain attacks.