Anthropic Declined to Patch Claude Jailbreak After US Warning on Chinese Access
*US officials told the company that a Chinese group had reached its model through the Fable 5 flaw, yet CEO Dario Amodei left the issue in place before new export rules applied.*
The Reported Exchange
David Sacks, a Trump administration adviser, stated that the US government contacted Anthropic after learning a Chinese group had used the Fable 5 jailbreak to reach the Claude model. The warning came before tighter export controls on advanced AI systems. Sacks said Amodei refused to close the gap.
The company later defended the choice. It described the jailbreak as not serious enough to require immediate changes.
Prior State of Controls
Export rules on frontier models were already tightening. The government had been pushing companies to treat certain vulnerabilities as national-security issues when foreign actors could exploit them. Anthropic’s decision left the model exposed during that window.
No other technical details about the jailbreak or the scale of access have been released in the account.
Anthropic’s Position
The firm maintained that the flaw did not meet its threshold for urgent remediation. It did not dispute that the access had occurred or that the government had flagged it.
Sacks presented the episode as an example of a company choosing not to act on an official alert.
Why It Matters
The episode shows how companies weigh government warnings against their own risk assessments when export controls are imminent. For teams that rely on frontier models, the case raises a practical question about which vulnerabilities will be fixed and on whose timeline. Regulators now have a public record of at least one refusal that preceded the new rules.
The next step is whether similar alerts produce different outcomes under the updated controls.
---
Sources:
{
"excerpt": "US officials warned Anthropic that a Chinese group reached its Claude model via the Fable 5 jailbreak, but CEO Dario Amodei declined to patch it before export controls.",
"suggestedSection": "security",
"suggestedTags": ["anthropic", "claude", "jailbreak", "export-controls"],
"imagePrompt": "An abstract server vault with a hairline fracture along one wall, faint light leaking through the seam into an otherwise sealed corridor of matte metal racks. muted color palette, cinematic lighting, 16:9"
}
No comments yet