$ cat briefs/daily/2026-07-04.md
2026-07-04
July 4, 2026 (Sat)
1 new standard · 1 government deal (late capture) · 0 papers
Top Stories
1. Anthropic proposes Cyber Jailbreak Severity (CJS) framework — first cross-lab AI jailbreak standard
- Alongside the July 1 Fable 5 global restoration, Anthropic published a companion framework assessing jailbreak risk on four axes: capability gain, breadth, ease of weaponization, discoverability (source)
- Five severity bands: CJS-0 (Informational) through CJS-4 (Critical); summed axis scores map to bands (CJS-1: 1–3.5, CJS-2: 4–6.5, CJS-3: 7–8.5, CJS-4: 9–10)
- Co-developed with Amazon, Microsoft, Google, and Glasswing partners — the three major cloud platforms and 150+ vetted organizations
- Simultaneously launched a HackerOne bug bounty (
hackerone.com/anthropic-cyber-jailbreak) specifically for Fable 5 cyber jailbreaks - New safety classifier deployed at restoration: >99% of jailbreak replication attempts blocked
- Why it matters: CJS is the first cross-lab AI jailbreak severity standard, modeled after CVSS (the software vulnerability scoring standard). If broadly adopted, it standardizes how labs, government agencies (CISA, NSA), and researchers communicate jailbreak risk — moving the field from ad-hoc incident-by-incident response to a shared measurement language. The four co-developers represent a majority of the world's frontier AI compute and two-thirds of the commercial AI market.
- → AI-Enabled Cyberattacks, Claude Fable 5 (source)
Paper Picks
None above threshold today. HuggingFace Daily Papers July 3–4: no Tier-1 org authors found in WebSearch fallback (direct fetch 403). ArXiv cs.AI/cs.CL July 3 shows no Tier-1 org papers above threshold.
Watch
-
California–Anthropic Claude deal (June 29, — Governor Newsom signed a first-of-its-kind partnership giving all CA state agencies, cities, and counties access to Claude at 50% discount + free workforce training. Reach: 238,000+ state employees, 500+ cities, 58 counties. Key signal: CDT + CalOES using Claude Security + Claude Code for state cyber defense (scanning, triaging, patching state code) — first documented Claude Code deployment for state-level government security. (CA Gov) (TechCrunch) → Anthropic
-
OpenAI GPT-5.6 Sol on Cerebras (announced July 1) — OpenAI will launch Sol on Cerebras wafer-scale hardware at up to 750 tokens/second (5–15× the 50–150 tok/s typical for frontier models). Initially limited to select customers. Broader Sol availability expected mid-July 2026. (aesopacademy) → GPT-5.6 Sol (and Terra, Luna)
-
Grok 5 (xAI) — still not released as of July 4. Q2 deadline (June 30) missed; July now the stated target with no specific date. Grok 4.5 remains in private beta at Tesla/SpaceX. → xAI
-
Gemini 3.5 Pro — still in limited Vertex AI enterprise preview; July GA target maintained but no firm date. Google has missed two delivery targets (June, then early July). → Gemini 3.5 Pro
New in Wiki
No new entity, model, concept, or paper pages created today.
Updates
- AI-Enabled Cyberattacks: Added CJS framework section (Anthropic + Amazon + Microsoft + Google; July 1 event); updated Key Events Timeline; updated sources and tags
- Claude Fable 5: Added "Cyber Safeguards" and "CJS Framework" subsections to Safety Architecture; added HackerOne reference
- Anthropic: Added two new Recent Activity entries: (1) CJS framework + HackerOne (July 1), (2) California Claude deal (June 29,; updated sources and tags