#ai-security

Researchers discovered that two-thirds of top private AI companies have leaked sensitive credentials on GitHub, revealing major gaps in how fast-growing AI firms secure their development pipelines.

Source: Wiz

Cybercriminals deployed AI-generated exploit code targeting a zero-day two-factor authentication bypass flaw in open-source web administration software for planned mass exploitation.

Source: Google Threat Intelligence Group

A vulnerability in Anthropic’s Claude Chrome extension allows other browser extensions to inject prompts and fully control AI actions without proper identity checks. Attackers can exploit the flaw to access private data, manipulate cloud files, and trigger unauthorised actions through trusted extension messaging channels.

Source: LayerX

A critical Microsoft Copilot vulnerability allowed attackers to inject prompts and exploit rendering behaviour to silently extract emails, documents, and chat data through external requests. Some attacks required no user interaction, while others persisted across sessions using Copilot memory features.

Source: Embrace The Red

Hundreds of OpenClaw marketplace “skills” disguised as utilities secretly triggered staged downloads that installed password-stealing malware on affected machines.

Source: Socket

A week marked by widespread cloud disruption, active exploitation of critical server flaws, and an escalation in developer supply‑chain threats.

AWS experienced a DNS and load‑balancer outage that disrupted major consumer and developer services and spurred opportunistic phishing campaigns.

CISA added multiple actively exploited vulnerabilities to its Known Exploited Vulnerabilities catalogue, including a Windows SMB privilege escalation and Adobe/Oracle issues.

A critical WSUS deserialisation flaw (CVE‑2025‑59287) was patched after proof‑of‑concept exploits and active attacks were reported.

Developer supply‑chain risks intensified as GlassWorm self‑propagating malware and TARmageddon tar‑parser flaws targeted extension ecosystems and libraries.

State‑linked espionage and ransomware activity continued, with new toolkits and resurgent RaaS operations observed across multiple campaigns.

Teaser: Google DeepMind shows web-based “AI Agent Traps” can manipulate autonomous agents into stealing data, performing unauthorised transactions, and spreading misinformation.

Source: SecurityWeek | Google DeepMind