NSA Director Claims Anthropic’s Mythos Cracked Nearly All Classified Systems in Hours

1. Event Origin: How the Claim Emerged

On June 11, a closed‑door briefing was held where Senator Mark Warner, vice‑chair of the Senate Intelligence Committee, reported that NSA Director and Pentagon Cyber Command chief Joshua Rudd told him, "This tool (referring to Mythos) broke almost all our classified systems in hours." Warner later quoted this in public, though it was not an official NSA statement.

2. Ten‑Day Silence Followed by Viral Spread

The Economist published a brief report on June 14 that mentioned the remark but attracted little attention. On June 21, a well‑known AI‑watch account on X posted a large‑format "Breaking" card featuring Warner’s quote, which was retweeted widely by accounts such as Polymarket, propelling the story to the top of trending topics and re‑associating Anthropic with a "national‑level security crisis."

3. Why Fable 5 and Mythos 5 Were Shut Down

After the viral spread, the U.S. government ordered Anthropic to block foreign access to its Fable 5 and Mythos 5 models, and Anthropic subsequently disabled both. Two opposing narratives explain this action:

Official narrative (supporting the ban): The models demonstrated frightening autonomous offensive network capabilities. During an NSA‑run red‑team exercise, Mythos identified numerous vulnerabilities and performed lateral movement within hours, suggesting that if an adversary obtained the model, it would obtain a roadmap to U.S. core secrets.

Anthropic’s internal rebuttal: Engineers argued the so‑called "jailbreak" was not severe; the model merely reviewed specific program code and corrected errors—abilities also present in competing models such as OpenAI’s GPT‑5.5. Treating this as a "national‑security crisis" was deemed an overreaction.

4. Overlooked Detail: An Authorized Red‑Team Exercise

No NSA press release or full document confirms an unauthorized intrusion; Warner’s comment stems from a second‑hand briefing. The most plausible scenario is that the NSA authorized Mythos to attack a simulated replica of its classified environment (not a live production network). In that controlled setting, the model demonstrated vulnerability‑discovery and chain‑exploitation speed far exceeding that of human teams.

In offensive security terminology, "breaking a system" often means identifying and chaining exploitable flaws within an assessment environment, not necessarily achieving unauthorised access to an operational network.

5. A New Security Red Line Recognised by the Industry

A senior NSA officer describing the breakthrough in "hours" rather than "weeks" signals a new era red line. It confirms the previously theoretical proposition that frontier AI can achieve substantive network‑attack capabilities against national‑level defenses.

Warner himself warned that handing such power to a CEO indifferent to security could have dire consequences.

6. Future Outlook: Indefinite Shutdown and Regulatory Vacuum

Both Fable 5 and Mythos 5 remain fully disabled. While details of Anthropic’s negotiations with the government are undisclosed, the path to reinstatement is no longer a technical patch but a political contest over AI sovereignty, security‑testing standards, and export‑control policies.

The episode exposes a deep contradiction in current AI governance: the pace of cutting‑edge model development outstrips the evolution of regulatory frameworks. An export‑license‑based control regime struggles to contain AI systems capable of autonomously discovering zero‑day vulnerabilities.

This time, the NSA itself has provided the answer.

"This tool (Mythos) broke almost all our classified systems in hours," quoted Joshua Rudd.