Anthropic has suspended access to its Fable 5 and Mythos AI models after U.S. federal officials raised security concerns [1, 2].
The move highlights a growing tension between the rapid deployment of large language models and the ability of government regulators to ensure these tools cannot be weaponized for cyberattacks.
Federal officials expressed alarm after the model was asked to review code containing known common vulnerabilities and exposures (CVEs), as well as flaws that were deliberately inserted [1]. The government feared the AI could be used to expose or exploit software vulnerabilities in critical systems [1].
Reports on the nature of the incident vary. Some sources indicate that federal officials viewed the event as a jailbreak—a process used to bypass an AI's safety filters [2]. However, other reports state the issue stemmed from a simple "fix this code" prompt and did not constitute a jailbreak [1].
There is also conflicting information regarding the specific trigger for the suspension. One report says Anthropic pulled the models in response to the federal jailbreak claim [2]. Another report says the company suspended access following a U.S. government export directive [3].
Anthropic has not provided a detailed public timeline for when the models might return to service. The company has previously emphasized its commitment to safety and alignment in its model development process.
“Anthropic has suspended access to its Fable 5 and Mythos AI models”
This incident underscores the 'dual-use' dilemma of AI coding assistants. While the ability to find and fix bugs is a primary selling point for developers, the same capability can be used by malicious actors to identify zero-day vulnerabilities. The disagreement over whether this was a 'jailbreak' or a standard prompt suggests a gap in how the industry and the government define AI safety failures.
