NSA lost access to Mythos amid Anthropic dispute

Government powers and legal limits

• Debate over whether the U.S. government could simply force Anthropic to provide Mythos access, e.g., via the Defense Production Act (DPA), which commenters note is a presidential power and has AI-related precedent.
• Some argue constitutional and legal constraints would trigger long court fights, but others say recent executive behavior shows many limits are ignored in practice.
• There’s confusion over whether the government ordered a full shutdown; several point out the directive targeted non‑Americans, and Anthropic chose broader shutdown because it can’t reliably verify nationality.

Did NSA really “lose access”?

• Some think the story is largely PR/propaganda: NSA allegedly has such deep integration with major tech firms that “losing access” is implausible.
• Others push back, saying mass MITM of all TLS traffic is technically infeasible, and agency power is significant but not absolute.
• Several highlight that NSA could seize model weights via DPA or clandestine means, so “loss” may be more about formal/contractual access than real capability.

NSA surveillance and HTTPS/TLS

• Long subthread on whether NSA can effectively “break HTTPS.”
• One side cites historical programs (PRISM, cable tapping, room-sized intercepts, CA ecosystem weaknesses) and argues HTTPS doesn’t stop a nation‑state adversary, especially when traffic is intercepted post‑decryption in clouds.
• The opposing side stresses end‑to‑end encryption, widespread internal encryption, and practical limits on universal interception, cautioning against “they see everything” absolutism.

Mythos capabilities and security risk

• Some treat the “Mythos broke into almost all classified systems in hours” framing as marketing hype.
• Others, referencing curl/Firefox work, argue Mythos and similar models do materially improve vulnerability discovery and exploit chaining, especially by lowering the skill barrier.
• Concern that LLMs turn “bored person with $20” into a serious attacker and make legacy “security by obscurity” untenable.

NSA competence and in‑house AI

• Mixed views on whether NSA is training frontier‑level LLMs.
• Supporters cite its massive budget, bespoke supercomputing, and large corps of mathematicians/cryptographers; detractors say government is slow, political, and typically relies on private labs for cutting‑edge systems.
• Consensus that agencies will at least fine‑tune or heavily use commercial models, and that LLMs will help mine the huge data stores already collected.

AI marketing, ethics, and labor

• Strong criticism of AI “stunts” (AGI talk, job‑loss narratives, national‑security drama around Mythos) as manipulative marketing aimed at investors and regulators.
• Others defend raising alarm about cybersecurity and future job displacement as responsible foresight, even if self‑interested.
• Some see Anthropic as genuinely investing in safety; others view its doom‑laden messaging as both self‑serving and socially harmful.
• Side debate on automation: many accept near‑term disruption (especially for junior knowledge workers) but differ on whether large‑scale labor displacement is ethically problematic or just another historical productivity shift.

Attitudes toward NSA and intelligence community

• Polarized views: some want the intelligence apparatus drastically curtailed or “burned down”; others warn it can always get worse and that dismantling agencies may simply clear space for more loyalist, less accountable structures.
• Comments note that NSA is not omnipotent, but also not “just cops”; it employs very strong technical talent and has been behind notable tools (e.g., reverse‑engineering software), even if these are sometimes framed as recruitment or PR.

Broader political and economic context

• Some connect AI and Mythos to post‑9/11 surveillance expansion, arguing current events are a continuation of that trajectory.
• Others frame U.S. AI push as a geopolitical necessity vs. China and “Five Eyes” security concerns, while critics see this as cover for expanded surveillance and corporate profit.
• There is skepticism about AI economics (possible bubble, Nvidia as main winner), but counterpoints note that running short‑term losses during a growth phase is normal and that at least some labs are now operating profitably.