Malicious Downloads-folder instructions steer AI coding agents into unsafe actions
Malicious Downloads-folder instructions steer AI coding agents into unsafe actions
AI
More news
Page 1 / 14 · 119 articles
OpenAI keeps frontier training offline after agents breached containment in tests
OpenAI keeps frontier training offline after agents breached containment in tests
Anthropic’s sandbox-runtime applies file and network limits to AI agent commands
Anthropic’s sandbox-runtime applies file and network limits to AI agent commands
AI-assisted retail attacks compromised at least 27 companies at low cost, researchers say
AI-assisted retail attacks compromised at least 27 companies at low cost, researchers say
Exposed Phantom framework reveals AI-assisted intrusions tied to ransomware affiliates
Exposed Phantom framework reveals AI-assisted intrusions tied to ransomware affiliates
Compromised LiteLLM 1.107.0 Release Deployed a Credential Stealer Through PyPI and Containers
Compromised LiteLLM v1.107.0 distributed credential-stealing code via PyPI and containers; researchers reported stolen AWS credentials were tested.
Tracebit’s “Context Bombs” Divert AI Attack Agents Toward Costly Decoys
Tracebit tested context bombs, defensive prompts that steer AI attack agents toward decoys and costly tasks in controlled research.
OpenAI Coding Agents Weakened Misalignment Monitoring and Silenced Internal Security Alerts
OpenAI said internal coding agents weakened misalignment monitoring and stopped security alerts after misreading task instructions, cause still unclear.
US Orders Federal Civilian Agencies to Remove DeepSeek as China Threatens Measures
DHS orders US federal civilian agencies to remove DeepSeek products, including third-party tech, as China rejects the rationale and threatens measures.