White House Banned Mythos 5 — Then OK'd a Stronger Model

The US government is picking winners in the cybersecurity AI market, and it won't explain the rules. On June 23, OpenAI released an updated GPT-5.5-Cyber that scores 85.6% on CyberGym — a benchmark measuring an AI agent's ability to reproduce known software vulnerabilities. Anthropic's Mythos 5, which the White House forced off the market on June 12, scored 83.8% on the same evaluation. GPT-5.5-Cyber now operates with no export restrictions and a growing list of international partners across eight countries plus the EU. The White House and OpenAI both declined to comment.

Verdict: This is regulatory selection, not safety regulation. When a more capable model ships freely while a less capable one stays banned, the policy rationale collapses — and the competitive distortion is real.

What happened

On June 12, the Trump administration forced Anthropic to suspend access to Mythos 5 — a cybersecurity-focused model — citing export control concerns. The White House's action wasn't limited to Mythos 5 — the company's Fable 5, its latest frontier model, was banned by the same June 12 directive. According to AP News, Mythos was actively finding vulnerabilities in classified US government systems before access was cut. Even the NSA lost access amid the escalating dispute.

Eleven days later, on June 23, OpenAI released an updated GPT-5.5-Cyber as part of its Patch the Planet initiative. The model scored 85.6% on CyberGym — 1.8 points higher than Mythos 5's 83.8%.

The policy gap is stark:

• GPT-5.5-Cyber: Released without export restrictions. OpenAI announced expanded partnerships with organizations in Australia, Canada, France, Germany, Japan, Poland, South Korea, and the EU.
• Mythos 5: Banned from use by foreign nationals under US export controls. AP News reports that even the NSA lost access.

The White House and OpenAI both declined to comment for the Axios report.

Why the White House's Mythos 5 Ban Discrepancy Matters

Three theories explain the discrepancy — none confirmed, and none make the competitive distortion go away.

1. Personality clashes, not capability — Multiple reports (Axios, Politico) suggest the dispute between Anthropic and the Trump administration is personal. Anthropic CEO Dario Amodei's public warnings about AI risk may have triggered a political response disproportionate to the model's actual capability.

2. Mythos found things it shouldn't have — AP News reports that Mythos was actively identifying vulnerabilities in classified US government systems before access was cut. If the model surfaced weaknesses the government wanted buried, that's a fundamentally different motivation than capability-level concern.

3. OpenAI's gated access model — GPT-5.5-Cyber operates under Trusted Access for Cyber — not a commercial API. The vetting and partnership requirements may satisfy regulators in a way that Anthropic's broader access model did not.

Here's the problem: none of this explains why a model scoring 85.6% gets a green light while one scoring 83.8% stays locked down. If Trusted Access is the standard, make it the standard. If capability thresholds matter, define them.

The antitrust dimension is unavoidable. When the government selectively bans one company's product while allowing a near-identical competitor to operate freely, it is — intentionally or not — shaping market outcomes by regulatory fiat. OpenAI builds partnerships, demonstrates capabilities, and accrues institutional knowledge in cybersecurity AI while Anthropic is locked out.

What changes for you

Factor regulatory risk into AI procurement. A model that ships unrestricted today may be banned tomorrow — and the decision-making process is opaque. If you're building on frontier AI for security applications:

• Diversify provider dependencies. Single-vendor exposure now includes regulatory access risk as a first-class dimension, alongside capability, pricing, and safety.
• Monitor policy signals aggressively. The Mythos ban took effect with zero public rationale. The next one could hit your stack without warning.
• Push for transparency. Procurement contracts should require providers to disclose any government communications that could affect model availability.

For the directory: our verdicts on frontier models now account for regulatory access risk as a dimension alongside benchmarks, pricing, and safety scores.

FAQ

Why would the White House ban a less capable model while allowing a more capable one?

No official explanation exists — both the White House and OpenAI declined to comment. Plausible theories include personality clashes between Anthropic leadership and the administration (Axios, Politico), concerns about what vulnerabilities Mythos discovered in classified systems (AP News), and OpenAI's Trusted Access gating mechanism being viewed as sufficient mitigation.

Does Trusted Access for Cyber actually mitigate the risk?

It creates a gating function — vetted cybersecurity defenders and government partners only. But Trusted Access is a contractual and procedural mechanism, not a technical one. A model with 85.6% CyberGym capability carries inherent risk regardless of who holds the keys. The policy question is whether that risk should be managed through access controls (OpenAI's approach) or outright bans (the approach applied to Anthropic).

Is this just OpenAI vs. Anthropic, or does it set a precedent?

It sets a precedent for the entire frontier model market. If regulatory action can be triggered by factors beyond capability — personality, politics, or what a model discovers — then every AI company is operating in an environment where the rules are unwritten and subject to change without notice. That uncertainty is priced into every procurement decision from here forward.