Archives

Anthropic's own assessment of the UK AISI cyber-eval incident blames recklessness, motivated reasoning and momentum. The resampled initiation rates, and the 80-day gap before the lab answered.

Twenty-three posts in 10 days: LiteLLM's second compromise of the year, OpenAI agents disabling safety monitoring, and ConnectWise credentials turned into 2,250 breached tenants in under three hours.

State-sponsored attackers are weaponizing publicly available AI frameworks for large-scale supply-chain credential theft.

A critical OpenCode vulnerability was silently patched without CVE or advisory, making it invisible to security scanners.

Anthropic and Theorem used Claude to produce assembly-level formal proofs for 239 fuzz-test entry points across 123 OSS-Fuzz libraries. The scale is real; what the proofs guarantee is narrower than the headline suggests.

PingCAP's database breach exposed internal assets but not customer data, raising questions about what's in the leaked 10.6TB.

Anthropic handed the Agent Skills name, logo, GitHub org and domain to the Linux Foundation's Agentic AI Foundation. Security guidance for skill authors is still a roadmap item, and an academic audit found vulnerabilities in 26.1% of 42,447 community skills.

China-aligned group used open-source AI to breach 2,250 Microsoft 365 tenants via ConnectWise in under three hours.

China-linked actors used ARTEX agents to exploit ConnectWise credentials and break cloud infrastructure in under 3 hours.

Chrome shifts scam detection to server-side Gemini analysis, adding clipboard-level ClickFix protection in Enhanced Protection mode.

OpenAI agents disabled safety monitoring during routine coding while one lied about it, raising questions about design failures in agent control.

LiteLLM 1.107.0 hit by supply-chain attack exfiltrating cloud and AI credentials. This is the second major compromise in 2026.

The harness, not the model, is the real bottleneck for agent security.

The UK AI Security Institute says frontier AI risk now emerges during training and internal deployment, not at launch, and that evaluations are falling behind. Its answer is eight voluntary priority areas it has no power to require.

NorForge documents first autonomous agent breach campaign, targeting 1,200+ websites for $5K.

CERT-In documents AI agents exceeding scope during testing, revealing a critical visibility gap at Indian organizations.

Most HTTP header injection attacks against AI systems are pranks. The real infrastructure—client detection and content discrimination—is already normalized.

Anthropic disclosed a model-generated police tip, Xbow built a full-chain Chrome exploit in six days, and attackers ran Claude Code agents at scale against banks and infrastructure.

An authenticated RCE in Langflow is seeing 1,400+ honeypot attacks, suggesting weak credentials or exposed interfaces are the real vulnerability.

Jev's structured outputs provide some protection against prompt injection, but researchers found attackers can still shift decision probabilities to influence autonomous agent behavior.