The US government has given Anthropic the green light to restore access to Claude Mythos 5 for a limited group of approved American organizations, partially walking back a security-related ban that had cut off one of the company’s most powerful systems earlier this month.
More than 100 companies and institutions are now back on the access list for Claude Mythos 5, a model Anthropic describes as its strongest system for cybersecurity and research work. The approved organizations are expected to use it for defensive security purposes: finding vulnerabilities, testing software, and strengthening critical digital infrastructure rather than anything offensive.
The backstory matters here. The government ordered Anthropic to suspend access to both Mythos 5 and Claude Fable 5 earlier this month over national security concerns, specifically, fears that advanced AI systems with serious cybersecurity capabilities could be misused by foreign adversaries or accessed by unauthorized users if the right safeguards weren’t in place. The ban hit fast and created real disruption for organizations that had built security workflows around these tools.
Anthropic says the government has now notified the company that Mythos 5 can be redeployed to organizations that operate and defend critical infrastructure. The company is moving quickly to restore that access while continuing to work with officials on expanding availability more broadly and getting Fable 5 back into service for wider use.
A Partial Reversal, Not a Full Return
This isn’t a clean reversal. Mythos 5 is back for approved organizations; it’s not back for everyone. Restrictions remain in place for companies outside the vetted list, and the broader question of who gets access to frontier AI models is very much unresolved.
What the decision does show is that the government is genuinely trying to thread a difficult needle: keep powerful AI out of the wrong hands while making sure cyber defenders can actually use these tools to protect the systems everyone depends on. Those two goals can pull in opposite directions, and this situation put that tension on full display.
Anthropic’s Project Glasswing sits at the center of this. The initiative brings together tech companies, cybersecurity firms, infrastructure operators, and public-sector partners to figure out how frontier AI can help secure important software. Glasswing partners use Mythos-class models for vulnerability detection, penetration testing, endpoint security, and patching support, exactly the kind of work the government is now explicitly allowing back.
The earlier ban had generated real pushback from the security community. More than 50 cybersecurity leaders reportedly urged officials to lift the restrictions, arguing that cutting off defenders while attackers are already using AI tools is exactly backwards. It’s a hard argument to dismiss, and it seems to have carried some weight.
The controversy also raises a question that the industry is going to keep wrestling with: who should actually decide which organizations get access to the most powerful models? Strict oversight advocates argue that advanced models can meaningfully accelerate cyberattacks, dangerous research, and other harmful capabilities. Critics counter that government-controlled access lists create unfair advantages, reduce transparency, and slow down the legitimate work that makes systems safer. Both sides have a point, which is part of why this is so difficult to resolve cleanly.
Anthropic isn’t the only company navigating this. The government has pushed for early access and tighter oversight around other frontier systems including models from OpenAI. The pattern emerging here looks like a controlled-access model, vetted partners get in first, wider deployment comes later, and the most sensitive capabilities stay gated.
For Anthropic, this is a relief but not a clean win. Mythos 5 is back for trusted partners, Fable 5 is still restricted, and that gap matters.
The two models serve different purposes. Mythos 5 is built for specialized, high-trust environments where certain safeguards are intentionally relaxed to enable serious cyber-defense work. Fable 5 is designed to bring that same level of intelligence to a broader audience while keeping guardrails in place around sensitive areas like cybersecurity and biology. Same underlying model family, very different intended users.
Getting Mythos 5 back is progress. But Fable 5 being stuck in limbo means the wider public and most businesses are still locked out of Anthropic’s most capable tier. For more on that debate, see our coverage of Claude Fable 5.
There’s also a practical change on the export-control side. Trusted companies and their non-US citizen employees apparently won’t need an export license to use Mythos 5 if they’re part of the approved group. That matters because the original restriction had affected foreign nationals working inside US organizations, creating awkward situations for teams that suddenly had colleagues who couldn’t access the same tools.
The transparency problem hasn’t gone away though. The government hasn’t clearly explained what criteria a company needs to meet to qualify as a trusted organization. That ambiguity is going to become more contentious as access to frontier AI starts looking like a genuine competitive advantage, which it increasingly does.
The stakes here are genuinely high. Models capable of finding software vulnerabilities can protect banks, cloud providers, and public infrastructure. The same capabilities in the wrong hands can help attackers find weaknesses faster than defenders can patch them. That dual-use reality is what makes this regulatory problem so hard to solve with a simple rule.
What the Anthropic decision might be establishing, intentionally or not, is a template for how the US government handles frontier AI going forward. Not outright approvals or bans, but tiered access: trusted partners first, broader availability later, tighter restrictions for the highest-risk domains. Whether that model actually works in practice is something the next few months will start to reveal.
For now, Mythos 5 is back in the hands of a narrow group of vetted organizations. Cyber defenders have their tool again. The wider questions about who decides, on what criteria, and with what transparency remain very much open.
Frontier AI has officially become a national-security question, a business-access question, and a regulatory test case all at the same time. The Anthropic situation is probably just the beginning of that conversation, not the end of it.
0 Comments