Anthropic embeds engineers at NSA for Mythos

// 4h agoNEWS

Anthropic embeds engineers at NSA for Mythos

A Financial Times report reveals that Anthropic has embedded engineers at the NSA to customize its restricted Claude Mythos Preview model for potential offensive cyber operations. Despite legal disputes with the Pentagon, the partnership is defended by advocates as a necessary step to counter AI-driven threats from hostile nations.

// ANALYSIS

This partnership highlights the thin line between defensive AI safety and government-sponsored offensive cyber warfare, showing that even safety-oriented labs will align with defense agencies when geopolitical stakes are raised.

–**Deep integration:** Embedding engineers directly within the NSA indicates a level of customization that goes far beyond standard commercial API access.
–**Geopolitical friction:** The collaboration is politically complex due to Anthropic's lawsuit against the Pentagon over supply-chain risk designations.
–**Offensive capability:** Claude Mythos Preview represents a leap in agentic security capabilities, with the autonomous ability to discover and exploit vulnerabilities.

// TAGS

anthropicclaude-mythos-previewnsasecuritysafetyoffensive-cybergeopolitics

DISCOVERED

4h ago

2026-06-05

PUBLISHED

7h ago

2026-06-04

RELEVANCE

8/ 10

AUTHOR

jawiggins

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

LAUNCH4m ago

Higgsfield has launched a new Figma plugin, integrating its powerful AI video and image generation capabilities directly into the design platform.

AI video generation platform Higgsfield has released a dedicated Figma plugin, bringing its suite of creative tools directly into the Figma and FigJam environment. The plugin allows designers to generate AI images and videos, remove backgrounds, relight frames, and apply color grading on their design canvas without needing to export assets from external applications.

NEWS21m ago

Theo highlights the split between the AGENTS.md standard and Anthropic's CLAUDE.md, explaining how Claude models require unique and detailed instruction steering.

In a post on X, content creator Theo discusses the emerging AGENTS.md repository context standard, noting Anthropic's decision to maintain its own CLAUDE.md format. While initially critical of the separation, Theo notes that Claude models require significantly more detailed, specific steering and "handholding" compared to other models like OpenAI's. Consequently, maintaining a separate, longer, and more customized CLAUDE.md file is often necessary for optimal agent performance.

UPDATE50m ago

OpenCode v1.16.0 integrates v2 sessions into core

OpenCode v1.16.0 brings v2 sessions directly into the native core API to enable native Effect-based embeddings. In addition to core API enhancements, this release introduces stronger workspace controls to both the App and Terminal User Interface (TUI), applies protocol fixes to model providers, and adds model-level pages in the Stats section for better tracking.