After dissing Anthropic for limiting Mythos, OpenAI restricts access to Cyber
AI companies’ claims that their latest cybersecurity-focused models are “too dangerous” to release are being met with deep skepticism, especially after OpenAI restricted access to its Cyber model soon after criticizing Anthropic for limiting Mythos. Commenters frame these moves as a mix of marketing, regulatory risk management, and compute economics rather than purely safety-driven decisions, noting unproven capabilities, rising access barriers, and governments’ growing role as key customers. The debate also touches on falling trust in leadership at major labs, the rapid progress of cheaper or local models, and whether any of the big players have a sustainable business model beyond hype and subsidies.
Hype, “Dangerous Models,” and Marketing
- Many see the “too dangerous to release” positioning of Mythos and Cyber as a marketing tactic: artificial scarcity, velvet ropes, and “my model is more dangerous than yours.”
- Some argue labs would release these models if it maximized revenue; withholding suggests either overhyped capabilities or genuine risk.
- Others think companies want to appear responsible and prepared in case their models are later linked to real-world cyberattacks.
Cybersecurity Capabilities and Verification
- Claims: current models have strong vulnerability-research capabilities and can find large numbers of bugs or vulnerabilities.
- Skepticism: lack of broad, trusted third‑party evaluation; some “benchmarks” are called anecdotal (e.g., tiny code samples).
- Some links and projects are cited as partial evidence that Mythos‑style capabilities are not uniquely beyond existing pay‑as‑you‑go models.
- Unclear whether Mythos is truly exceptional or just good marketing around incremental improvements.
Economics, Pricing, and Compute
- Discussion of DeepSeek V4 pricing being dramatically lower than OpenAI’s models; some suspect state subsidy, others note US tech has long been effectively subsidized too.
- Debate over whether inference is actually being subsidized: some say nobody is profitable at scale; others insist third‑party hosts and major providers have healthy per‑token margins.
- Compute scarcity and long lead times are seen as a major strategic factor; pre‑locking capacity may matter more than model quality.
OpenAI, Leadership, and Trust
- Strong distrust expressed toward OpenAI’s leadership, citing past reversals (e.g., RAM capacity rhetoric) and the CEO’s reputation for ruthlessness.
- Some note employees and media heavily backed leadership during prior board drama, suggesting internal loyalty but also a susceptibility to narrative management.
Safety Filters and Cyber Programs
- Users report more refusals on legitimate defensive security tasks and describe the Trusted Access Cyber program and its outsourced verification as clumsy.
- Debate on whether it’s technically possible to reliably distinguish offense from defense via text alone; some say in principle yes, others point to current tools as evidence of practical failure.
Local and Open Models vs Frontier
- Several argue local models are now “good enough” for many tasks and lag frontier models by only 6–12 months, undermining big labs’ moats.
- Examples of strong local models and new architectures are mentioned, though some users report reliability and context‑length issues.