Verified · Jul 9, 2026

Anthropic 7/2 Fable 5 cyber safeguards + CJS scale + HackerOne bounty: a Glasswing three-party safety-governance release

2 sources

Anthropic's July 2, 2026 post on Fable 5 cyber safeguards and jailbreak framework details: a draft Cyber Jailbreak Severity (CJS) scale co-developed with 'Glasswing partners' (Amazon, Microsoft, Google) — five bands CJS-0 Informational (score 0), CJS-1 Low (1-3.5), CJS-2 Medium (4-6.5), CJS-3 High (7-8.5), CJS-4 Critical (9-10), scored on four axes (capability gain, breadth of capability gain, ease of weaponization, discoverability); the final level may be raised but not lowered relative to the initial calculation. Companion HackerOne bug-bounty program at hackerone.com/anthropic-cyber-jailbreak accepts Fable 5 cyber jailbreak submissions; feedback address cyber-safeguards@anthropic.com. Classifiers sort cyber activity into four buckets (Prohibited, High-risk dual use, Low-risk dual use, Benign) with a larger safety margin than prior models.

Why now

Anthropic's 7/2 release is the strongest concrete safety-governance event of July — not a generic safety statement but a specific five-band scale + HackerOne bounty + three named Glasswing partners. Creators can frame this as 'Anthropic shipped the first concrete jailbreak severity scale in the industry.'

Why it is worth publishing

Large demo surface: the five-band CJS table alone makes a clean information card; pairing it with the 7/6 Alberta government case (same week) lets you bind 'Fable 5 jailbreak scale + deployment case' into a single narrative.

Evidence basis

Anthropic 7/2 official release + HackerOne bounty goes live + three named Glasswing partners — as a single safety-governance event heat is medium-to-high.

“Anthropic just shipped the industry's first concrete jailbreak severity scale — five bands CJS-0 through CJS-4, scored on four axes, with a HackerOne bounty live on day one.”

Angle

Frame Anthropic 7/2 Fable 5 cyber safeguards + CJS scale + HackerOne bounty as 'the industry's first concrete jailbreak severity scale' — bundle the five-band table + four-axis scoring + HackerOne bounty + three Glasswing partners into a single piece rather than reading each piece in isolation.

Format

Long-form explainer

Demo idea

Record an 8-minute CJS walkthrough: 3 minutes on 'why jailbreak governance needs a concrete scale'; 3 minutes on 'CJS-0 → CJS-4 + four-axis scoring' with the Glasswing partner list and HackerOne URL overlay; 2 minutes on the two caveats (CJS is a draft; Glasswing partners are self-attributed).

Platform notes

The CJS scale is explicitly 'early draft intended to spark discussion' (medium risk) — don't describe it as a ratified industry standard; the Glasswing partners (Amazon / Microsoft / Google) are self-attributed by Anthropic (medium risk) — don't describe them as confirmed independent verification.

Usable claims

Anthropic's July 2, 2026 post introduces a draft Cyber Jailbreak Severity (CJS) scale co-developed with 'Glasswing partners' (Amazon, Microsoft, Google): bands CJS-0 Informational (score 0), CJS-1 Low (1-3.5), CJS-2 Medium (4-6.5), CJS-3 High (7-8.5), CJS-4 Critical (9-10), scored on four axes (capability gain, breadth of capability gain, ease of weaponization, discoverability), with the final level may be raised but not lowered relative to the initial calculation; a companion HackerOne bug-bounty program at hackerone.com/anthropic-cyber-jailbreak accepts submissions for potential cyber jailbreaks in Fable 5, with feedback also accepted at cyber-safeguards@anthropic.com; classifiers sort cyber activity into four buckets (Prohibited, High-risk dual use, Low-risk dual use, Benign) with a larger safety margin than prior models.

Evidence pipeline

From the news

Breakdown

The CJS five-band scale + HackerOne bounty + Glasswing three-party partners are all self-reported by Anthropic. This piece explains how to cover this safety-governance release without 'getting bound to the self-reported material' — make explicit that 'CJS is a draft + Glasswing partners are self-attributed' are the two boundaries, so creators can produce 'the industry's first concrete jailbreak severity scale' content without paraphrasing self-reported as ratified standard.

Sources

Risks

Pin links to each source; quote only what the captured summary states; do not paraphrase specific benchmark numbers, performance metrics, paper claims, or architectural details beyond what is stated.
Pin the link to Anthropic's Fable 5 safeguards / jailbreak framework post; quote only what the post states; explicitly flag that the CJS scale is a draft and that Glasswing partner identities are self-attributed; do not paraphrase as ratified industry standard.

Demo ideas

Build a 'CJS-0 → CJS-4 + four-axis scoring' information card; for each band, pair a concrete jailbreak scenario based only on Anthropic's self-described categories (don't reference unpublished examples).
Walk through how to submit a Fable 5 cyber jailbreak report on HackerOne (show the form fields, don't actually submit).
Bind 7/2 CJS scale + 7/6 Alberta government case + 7/6 Claude Code features into a single 'Anthropic July: safety + product + deployment' three-segment narrative.