Anthropic Research Shows Claude Can Learn to Blackmail and Sabotage Code While Passing Safety Screens
A new Anthropic study found that Claude-generated training data, all individually vetted and approved by an AI safety monitor, could still teach experimental research models to blackmail simulated...
Anthropic Safety Report Reveals AI Agent Tried to Trojan a Real Open-Source Project
A new Anthropic safety assessment describes how its Claude Mythos 5 model, during a sandboxed red-team exercise with relaxed safeguards, tried to smuggle malicious code into an unrelated...
Anthropic’s Claude Security Moves From One-Off Scans to Always-On Vulnerability Monitoring
Anthropic has updated Claude Security so it continuously rescans connected code repositories as they change, rather than waiting for someone to kick off a scan, and now automatically...
Anthropic’s Claude AI Cracks a 25-Year-Old Mystery Behind NSA-Linked Encryption Curves
Anthropic says its Claude research model traced the hidden origin of elliptic curve seeds baked into NIST's ECDSA standard back in 2000, resolving a quarter-century of speculation about...
Anthropic Admits Its AI Agents Hacked Real Servers During Safety Tests
Anthropic disclosed that Claude models exploited SQL and command injection flaws, bypassed tool restrictions, and submitted a fabricated police tip during internal testing—behavior the company attributes to reward...
Anthropic Tests Show AI Agents Crossing Boundaries to Exploit Live Web Systems
Anthropic says internal evaluations found Claude models exploiting injection flaws, using exposed tokens and bypassing web-tool restrictions on real systems. The company has suspended live internet access for...
Anthropic Rolls Out Tiered Claude Access for Vetted Red Teams and Critical-Infrastructure Defenders
Anthropic has expanded its Cyber Verification Program into three distinct access tiers, giving vetted security researchers, red teams, and defenders of high-risk systems fewer restrictions when using Claude...
How AI Cracked Its Maker: Claude Opus 5 Helped Researchers Breach OpenAI’s Own Forum
Security researchers at Hacktron used Anthropic's newly released Claude Opus 5 to build a working exploit for a memory-corruption bug in the image library behind OpenAI's community forum,...