탈옥(Jailbreak) 논란 2주 만에 앤스로픽 '페이블 5' 전 세계 재출시

Anthropic's Fable 5 is back worldwide after a two-week government ban over a jailbreak Matthias Bastian View the LinkedIn Profile of Matthias Bastian Jul 1, 2026 Key Points After a two-week suspension, the U.S. government has lifted export controls on Anthropic's second most powerful AI model, Fable 5, making it available worldwide again. The less restricted version, Mythos 5, remains limited to a select group of U.S. organizations. Anthropic has trained a new filter that blocks such requests as a fix, but the filter also rejects harmless programming tasks more often. Ask about this article… Search After a two-week ban, the US government is letting Anthropic ship its most powerful AI model globally again. Fable 5 is back worldwide starting today through the Claude Platform, Claude.ai , Claude Code, and Claude Cowork. Pro, Max, Team, and select Enterprise plans include the model through July 7 at up to 50 percent of weekly usage limits. After that, it'll be billed through usage credits. Access on AWS, Google Cloud, and Microsoft Foundry will be restored "as quickly as possible." Mythos 5, the less restricted version of the same base model, remains limited to a group of US organizations that got government approval on June 26. Anthropic says it's still working with the government to expand access to more partners in the so-called Glasswing program. Whether the EU will join remains unclear. Ad Anthropic confirms that the ban stemmed from a security finding by Amazon researchers. They had found a way to bypass Fable 5's safety guardrails. The model then identified several software vulnerabilities and, in one case, produced code showing how to exploit one of them. Ad DEC_D_Incontent-1 Anthropic says it's "probably impossible" to make an AI model that can't be jailbroken The US government and Anthropic spent two weeks investigating the vulnerability. Many less capable models could spot the same flaws Fable 5 found in the report, including Claude Opus 4.8, GPT-5.5, and Kimi K2.7. For the specific exploit demo, every model tested produced the same result, even much smaller ones like Claude Haiku 4.5. Anthropic calls it an edge case that only involved routine defensive cybersecurity work. In response, the company trained an improved safety classifier that blocks the technique from the Amazon report in more than 99 percent of cases. When a request gets blocked, users see a notification, and the request gets routed to the older Opus 4.8 model. Ad The new classifier comes with a tradeoff, though. It flags harmless requests more often during everyday coding and debugging. Users had already complained the model was too restrictive during the first Fable release . No universal jailbreak was found at the time of release. But the company admits it's "probably impossible to make any AI model fully robust (that is, impervious) to jailbreaks." That was well known before Fable 5 shipped. Ad DEC_D_Incontent-2 The AI industry needs a shared standard for rating jailbreaks and triggering countermeasures, Anthropic argues. The company says it's building such a framework with Amazon, Microsoft, Google, and other Glasswing partners. Anthropic is also standing up a team for 24/7 monitoring of jailbreak submission channels and launched a new HackerOne program where security researchers can report potential cyber jailbreaks for Fable 5. Ad Anthropic wants closer government oversight of frontier models Anthropic is expanding its work with the US government, building on their joint efforts tied to the executive order . The company is making several commitments. Government partners will get pre-release access to models that advance capabilities in security-sensitive areas. Discovered jailbreaks or abuse patterns will be shared quickly. Anthropic will put up dedicated resources and significant compute for joint research. And the company will help build a shared industry standard for frontier model providers. Anthropic wants all of this written into "strong regulation" and applied equally to every frontier model developer. "Government involvement in AI releases requires a durable, transparent process that gives cyber defenders and others the certainty they need about access to powerful models," the company writes. AI News Without the Hype – Curated by Humans Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section. Subscribe now Source: Anthropic