NSA director: 'Mythos "broke into almost all of our classified systems in hours"
Claims that Anthropic’s Mythos AI model “broke into almost all” NSA classified systems within hours have prompted intense scrutiny of both the underlying security and the way the incident was communicated. Commenters note that a later clarification framed this as a red‑team exercise on internal networks, not an autonomous external breach, and argue the soundbite likely exaggerates or oversimplifies what actually happened. The exchange broadens into concerns about systemic software vulnerabilities, the rapid spread of Mythos‑level capabilities to other or open‑source models, and the use of dramatic cybersecurity narratives to justify political decisions and potential AI restrictions.
Context of the Mythos/NSA Claim
- Thread centers on a reported remark that Mythos “broke into almost all of our classified systems in hours.”
- Several commenters note this is second-hand (senator quoting NSA director) and likely compressed or misinterpreted testimony.
- Later update (via linked post) says the journalist clarified: Mythos was used by internal red-team pentesters with high-side access to speed up compromising systems, not an autonomous, external breach.
- Many see the original framing as exaggerated or misleading, possibly to justify new restrictions on Anthropic or AI models more broadly.
How Mythos Was Likely Used
- Common assumption: Mythos was placed “inside the perimeter” or on the same internal networks, assisting humans rather than acting alone on the open internet.
- Some suggest “broke into” may conflate any successful exploit (including simple crashes) with serious remote code execution.
- Several argue that if Mythos works this well, other top models (commercial or open) could likely achieve similar results given time and access.
Security Posture of Government Systems
- Multiple comments stress this reflects more on NSA/DoD security than on Mythos’ uniqueness.
- Claims that high-side systems rely heavily on air-gapping but run old, COTS, poorly maintained software with significant tech debt.
- View that many vulnerabilities are already known or trivial; AI mainly makes exploitation cheaper and faster.
- Some note the long-standing availability of more robust designs (capability-based microkernels, seL4, KeyKOS, strong isolation), arguing secure general-purpose computing has been “solved” in principle but not adopted.
Open-Source and AI Proliferation Concerns
- Debate over how far open-source models lag behind state-of-the-art (estimates from a few months to a year+).
- Some argue nonproliferation controls on powerful models are unrealistic; adversaries and open projects will continue regardless of US bans.
- Others say even if open models reach Mythos-like capability, defenders will have had more time to harden systems.
Hype, Propaganda, and Skepticism
- Significant skepticism that the US possesses an AI able to autonomously crack “all classified systems,” comparing this to prior intelligence-community exaggerations.
- Some view the narrative as effective propaganda or marketing for both Cyber Command and Anthropic.
- General sentiment: AI-assisted pentesting is serious, but the headline claim is likely an overhyped distortion.