This closes the loop on the export ban story we've tracked all week. Fable 5 is back for everyone starting July 1, but the real story is what Anthropic revealed about the incident: the "bypass" worked on nearly every model tested, not just Fable 5, and it's now proposing an industry-wide standard for scoring jailbreak severity.
- Claude Fable 5 becomes available globally again starting July 1, after the US government lifted export controls imposed on June 12.
- Anthropic’s investigation found the reported “bypass” could be replicated by nearly every model it tested, including Opus 4.8, GPT-5.5, and Kimi K2.7, and did not involve any unique Mythos-level capability.
- A new safety classifier blocks the specific reported technique in over 99% of cases. Blocked requests are automatically rerouted to Opus 4.8.
- Anthropic, Amazon, Microsoft, and Google are jointly drafting a four-criteria industry framework to score AI jailbreak severity: capability gain, breadth, ease of weaponization, and discoverability.
- Anthropic is expanding government collaboration, including pre-release access for national-security-relevant models and participation in the interagency cybersecurity clearinghouse from the June 2 executive order.
What Happened
Anthropic announced that Claude Fable 5 will be available globally again starting July 1, after the US government lifted the export controls it applied on June 12. For Pro, Max, Team, and select Enterprise plans, Fable 5 will count against up to 50% of weekly usage limits through July 7, after which it moves to usage credits.
The export ban followed a report from Amazon researchers describing a way to bypass Fable 5’s safeguards by prompting it to identify software vulnerabilities, in one case producing code demonstrating an exploit. Anthropic’s investigation found this was not a unique risk to Fable 5.
Nearly every model it tested, including Opus 4.8, GPT-5.5, and Kimi K2.7, could identify the same vulnerabilities, and every model tested, down to Claude Haiku 4.5, could produce the same exploit demonstration once prompted the same way.
The technique did not expose any Mythos-level capability and only touched a borderline case in Fable 5’s safeguards involving routine defensive cybersecurity work.
Anthropic trained an improved safety classifier that blocks the specific reported technique in over 99% of cases, with blocked requests automatically rerouted to Opus 4.8.
Researchers from the Commerce Department’s Center for AI Standards and Innovation tested both the old and new safeguards and confirmed them as extraordinarily strong.
Anthropic also confirmed Mythos 5 access has been restored for a set of US organizations following the government’s June 26 approval, with work ongoing to expand it to the broader Glasswing partner network.
Anthropic, alongside Amazon, Microsoft, Google, and other Glasswing partners, is now drafting a consensus industry framework to score jailbreak severity on four criteria: how far beyond existing tools a jailbreak takes the user, how many distinct offensive tasks it works for, how easy it is to turn into a real attack, and how discoverable the technique is.
Anthropic is also expanding its government collaboration, including pre-release access for national-security-relevant models, rapid safeguard information sharing, dedicated joint research teams, and participation in the cybersecurity clearinghouse established under the June 2 executive order.
Why It Matters
The most important detail in this whole episode is easy to miss: the vulnerability that triggered an 18-day global export ban was not unique to Fable 5 at all. Nearly every frontier model, including several from competing labs, could reproduce the same behavior. That reframes the entire incident.
This was not really about one lab’s model being uniquely dangerous. It was about the industry lacking a shared standard for judging how serious a given jailbreak actually is before governments act on it. The proposed four-criteria framework, if adopted industry-wide, would give both AI labs and regulators a consistent way to triage findings instead of reacting to individual reports in isolation.
For SaaS teams that depend on frontier AI models, this incident is a preview of how the August 1 government access framework will likely operate in practice: fast government action on incomplete information, followed by weeks of technical investigation that often reveals the risk was narrower than the initial response suggested.
Anthropic’s own new classifier trades false positives for safety, meaning legitimate coding and debugging requests will get blocked more often going forward. Relve, an AI trends intelligence platform, is tracking whether the proposed jailbreak severity framework gets adopted broadly enough to reduce the frequency of these disruptive, ban-first-investigate-later cycles.
Claude Fable 5 will be available again globally tomorrow.
After a series of productive conversations with the US government, we’re redeploying the model with a new set of classifiers to target and block more cybersecurity tasks. In the near term, some routine tasks like coding…
— Anthropic (@AnthropicAI) July 1, 2026
Bottom Line
Watch whether Amazon, Microsoft, and Google formally adopt the proposed jailbreak severity framework in the coming weeks. If a shared standard actually takes hold across frontier labs, it could meaningfully reduce the odds of another surprise 18-day model suspension triggered by an incomplete initial report.
For SaaS teams running production workloads on Fable 5, expect more false-positive refusals on borderline coding and cybersecurity tasks going forward. Anthropic has explicitly traded some legitimate request friction for a much larger safety margin. Budget for that friction in your workflows rather than treating every refusal as a bug.
