Breaking

Anthropic to Draft Shared Framework With Amazon, Microsoft, Google to Rate AI Jailbreaks

July 1, 2026 at 08:47 EDT

Security
Policy & Regulation
Foundation Models

On July 1, 2026, Anthropic announced it has begun drafting a "consensus framework" with Amazon, Microsoft and Google to objectively assess the severity of jailbreaks against AI models. At the same time, it disclosed the global redeployment of its latest model, Claude Fable 5, which had been suspended under export restrictions.

June 30, 2026 · Anthropic

A Shared Yardstick for AI Jailbreaks

Anthropic is drafting the first cross-company framework to measure the severity of AI safety bypasses — built jointly with Amazon, Microsoft and Google — after a June export-restriction episode exposed the absence of any shared standard.

assessment axes in the draft severity framework

99%+

block rate of the new classifier against the reported technique

companies co-authoring the standard

Cyber capability benchmark — CyberGym score

83.1%

Mythos Preview

66.6%

Opus 4.6

The four severity axes

Severity as a spectrum — not a binary of "broken or not."

Capability gain

Already reachable with existing tools — or greatly accelerating experts?

Breadth of gain

How many types of attack tasks one jailbreak enables.

Ease of weaponization

A single prompt — or many retries and real skill required?

Discoverability

Specialist knowledge needed — or already widely known online?

How the episode unfolded

Jun 12

Export restrictions applied to Fable 5 & Mythos 5

→

Response

New classifier trained, blocking the technique >99%

→

Jul 1

Fable 5 resumes global use; Mythos 5 limited to some U.S. orgs

Welcomed

Progress on governance; the four-axis, severity-based approach is seen as practical and overdue.

Concerns

The classifier can block routine coding and debugging — raising false positives and pushing some users back to Opus 4.8.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.