Anthropic model identifies vulnerabilities in classified US government systems

Anthropic's Mythos model identified security flaws within sensitive and classified government infrastructure during safety testing to evaluate model risks and capabilities, according to a US official.

The evaluation demonstrates that frontier models can uncover security flaws in classified networks during safety testing, showing their capacity to audit protected infrastructure.

  • The Mythos model detected flaws in secure, non-public systems.
  • The findings came from collaborative safety evaluations between Anthropic and US federal agencies.
  • The testing shows large language models can carry out complex security auditing tasks on sensitive networks.
Generative AI Sovereign AI AI Apps & Platforms
All AI news

More AI news

Models

AMD experimental figures report automated porting of 51 models to Strix Halo

AMD has released experimental figures showing an automated AI process ported 51 models to its Strix Halo platform in four days. The company reports that reading long text is faster on the hardware, with processing 32,000 tokens reaching a median speedup of 1.5 times. Maximum processing performance reportedly reached approximately 2.1 times faster in test conditions.

Industry

WhiteMatter raises 20 million dollars in series A funding for AI guardian agents

WhiteMatter announced a 20 million dollar Series A funding round on 10 November 2026. The investment will support the development of software designed to supervise other AI agents.

Industry

How enterprises can secure autonomous AI agent access

In an interview published by CIOL, Akkasha Sultan addressed how enterprises can evaluate and manage autonomous access for AI agents. Sultan explained that organisations can secure these deployments by implementing technical controls, specifically identity verification and delegated permissions for automated processes.

Build it with us, or have us check yours

In 15 minutes we talk through what you want to build, who builds it and what it needs to go live.

Talk directly with our AI specialists

15 minutes, no obligation
No sales pressure
Prototype in 7 days