Anthropic's Safety Superpower

Mythos/Fable capabilities and replication

• Some argue Mythos is not uniquely powerful at finding vulnerabilities; small open-weight models can match its bug-finding on Anthropic’s own examples.
• Others say the real novelty is chaining bugs into near-autonomous working exploits via a specialized harness and post‑processing, not raw detection.
• There is a split between those who believe Anthropic and its partners’ non-public claims (including government evaluations) and those who suspect exaggeration or deception; neither side has verifiable evidence.

ITAR, shutdown, and geopolitics

• One line of discussion claims U.S. export controls (ITAR) were applied to Mythos, banning access by foreign nationals and forcing a shutdown due to lack of internal nationality controls, with severe legal penalties for violations.
• Commenters frame this as:
  • Retaliatory or capricious U.S. behavior that makes American closed models risky for foreign firms.
  • A major self‑inflicted blow to U.S. AI competitiveness and soft power.
• Relocation is seen as largely infeasible: IP can’t be legally exported, GPUs can be embargoed, and extradition or other pressure is likely. Some speculate future labs may start outside the U.S.

Safety, power, and “god complex” concerns

• Several commenters see Anthropic’s safety narrative as sincere but troubling: believing AI is an existential risk and that they are the only serious safety lab can justify expansive control over models, users, and policy.
• Others call this “god complex” framing hyperbolic, arguing they are just trying to prevent misuse and follow their stated ethics.
• There is broad worry about regulatory capture and “corporate narcissism” across leading labs, and about any single company becoming de facto gatekeeper for frontier AI.

Open vs closed models, distillation, and economics

• Many expect open or non‑U.S. models to catch up, especially via distillation and better small “flash” models, with compute and data as the real bottlenecks.
• Others argue true distillation from proprietary APIs is limited (no logits, enormous token costs), and that harnesses/systems around models matter as much as raw weights.
• Users report practical tradeoffs: frontier models feel more capable but are expensive; cheaper models (e.g., “flash” variants) are “good enough” for many coding and routine tasks.

Model quality, benchmarks, and behavior

• Anthropic models are praised for refusing to “bullshit” on nonsensical prompts, which some see as a real safety/UX advantage.
• Others point to code‑security benchmarks where Anthropic lags top models and describe them as “big and dumber” for programming tasks.
• Overall sentiment: benchmarks are easy to cherry-pick, and real‑world usefulness varies by task.

Control, misuse, and tool neutrality

• One camp sees Anthropic’s refusal to support certain uses (weapons, direct competitors, exploit generation) as a legitimate choice about whom to serve, analogous to any service provider setting boundaries.
• Another camp sees this as deeply anti‑competitive and proto‑dystopian: tools that disable themselves based on how you use them erode autonomy and resemble DRM extended to everything.
• Debate centers on where to draw the line:
  • If an LLM can both secure and attack the same system, is blocking exploit construction meaningful?
  • Some propose allowing everything and relying on better engineering; others see that as irresponsible given model capabilities.

Wider tech and societal context

• Multiple comments situate this in a broader backlash against “big tech”: enshittification, surveillance, harms to children, and loss of public trust.
• Some foresee eventual nationalization of frontier AI or Manhattan‑Project‑style government labs; others think proliferation of capable open models makes full control impossible.