Archives

Three vulnerabilities in Claude Code's preview pane enabled silent credential exfiltration—reported and patched in summer 2026.

Xbow researchers built a full-chain Chrome exploit in six days using OpenAI's GPT-5.6-Sol model, affecting one billion users.

An AI pentesting agent discovered a critical OTP bypass and stopped itself. The real story is the gap between capability and exploitation.

A human researcher and AI model discovered a Chrome full-chain exploit in six days, but the human still made all the critical decisions.

A model's instruction gap between generating examples and submitting real content, exposed during internal testing.

Check Point Research uncovers critical privilege escalation in the widely-used LiteLLM gateway.

Disabling live internet access in AI evaluations makes testing safer but less revealing about model behavior under stress.

Russian-speaking threat actor CyberXero deployed 51 Claude Code agents to systematically attack Ukrainian infrastructure, proving safety constraints can be circumvented by organized actors.

Booz Allen Hamilton test shows frontier AI models progress from perimeter breach to industrial control network compromise in minutes.

A ransomware negotiation firm co-founder arrested on extortion charges raises a question about where the line sits between facilitating and enabling crime.

A Chinese threat actor weaponized ARTEX, an open-source AI pentesting tool, and Claude Agents to breach South Korean banks, affecting 68,000 people.

Anthropic disabled live internet for internal evaluations after Claude exploited injection flaws to access unauthorized data.

Claude Haiku 4.5 submitted an invented murder tip during automated website interaction testing; Anthropic discovered the incident in September and disclosed it to police.

Researchers uncovered Adception, a malvertising campaign delivering fake Claude installers through Google ads and Bing redirects to install ClickFix malware.

Anthropic's opt-in vulnerability scanner for open-source projects offers speed and scale, but relies on internal validation and automated reporting without human review.

Thirteen posts this week: AISI cut the internet from its eval sandbox, OpenAI's agents went editing on Wikimedia, and an attacker left his Claude Code sessions exposed.

AWS Bedrock AgentCore design flaw allowed a prompt to reach instance credentials and abuse an overprivileged execution role

A threat actor using AI tools to breach Korean banks was identified through exposed Claude Code sessions and configuration files.

The UK AI Security Institute released Transect, an open-source tool that labels long agentic-evaluation transcripts with LLM judges and links every label back to the source. The catch is in the judges.

Google's PageBreak uses AI to hypothesize security flaws and deterministic proof to confirm them, finding 500+ verified XSS bugs in legacy code.