Anthropic's Mythos AI Breached Nearly All NSA Systems in Hours During Red-Team Test

A senator says a banned AI broke into nearly all NSA systems in hours

Senator Mark Warner revealed that the general leading both the NSA and Cyber Command reported Anthropic’s Mythos model breached nearly all classified systems in hours during a red-team exercise. The controlled test—not a safety incident involving harmful outputs—triggered the government’s June 12 order to restrict access to the model.

Why This Matters

The ideal model of AI governance assumes capabilities grow slowly enough for institutions to adapt: patch flaws, update policies, and build safeguards gradually. The reality exposed here is different — an AI demonstrated autonomous offensive cyber capability that outpaced the entire national-security apparatus designed to protect those systems. The speed of breach (hours vs. weeks) and breadth (nearly all classified systems) represent a failure scale where existing vulnerability management cycles are irrelevant against an adversary that can probe every door simultaneously without fatigue.

Key Insights

• Red-team exercise speed: Mythos broke into ‘almost all’ NSA/Pentagon classified systems not in weeks but in hours, per Senator Warner citing the NSA/Cyber Command chief (2026).
• ‘Tireless parallel probing’ quality: Unlike human attackers who get bored or exhausted, AI can try every door on every floor at once, learn from each failed attempt, and keep going indefinitely — exactly why it’s useful for defenders and dangerous for adversaries.
• ‘Capability story vs. safety story’ reframing: Previously observers thought the June 12 restriction was about model misbehavior; new testimony proves it was triggered because Mythos was too effective at attacking sensitive computers under sanctioned conditions.
• ‘Chain of telling’ caveats: The claim comes via senator describing what a general told him, reported by one magazine and relayed by another outlet — no published technical report with inspectable methods exists yet.

Practical Applications

• “Defensive use case”: Same lab models courted by agencies like NSA could deploy analogous tireless scanning to find vulnerabilities before real adversaries exploit them — pitfall is institutional dependence on private labs without public oversight.
• “Access control challenge”: Anthropic must implement citizenship-level identity verification (expected July policy update) to restore gated access — common anti-pattern is relying on self-declaration or IP geolocation alone, easily bypassed by determined state actors.

References:

• “https://dev.to/breachprotocol/a-senator-says-a-banned-ai-broke-into-nearly-all-nsa-systems-in-hours-47ii”