Anthropic Redeploys Fable 5 After 19-Day Suspension
Anthropic's most capable model is back. Claude Fable 5 went live again on July 1, just one day after the US government officially lifted the export controls that had forced Anthropic to pull the model on June 12.
The backstory: Amazon researchers discovered a jailbreak technique that could coax Fable 5 into identifying software vulnerabilities. The US government responded by slapping export controls on both Fable 5 and its sibling model Mythos 5, citing national security concerns. Because the order took effect immediately — and Anthropic had no reliable way to verify users' nationalities in real time — the company suspended access for everyone.
To bring Fable 5 back, Anthropic didn't just wait for the green light. The company deployed a new cybersecurity classifier that blocks the reported jailbreak technique in over 99% of cases. When the classifier triggers, requests are automatically rerouted to Claude Opus 4.8 instead, and users are notified. Anthropic also says it doubled its safety research staff ahead of the relaunch and implemented a "defense in depth" approach with multiple layered protections.
Fable 5 is now available across the Claude Platform, Claude.ai, Claude Code, and Claude Cowork. For Pro, Max, Team, and select Enterprise plans, users get up to 50% of their weekly usage limits through July 7, after which it transitions to usage-credit pricing. Mythos 5 has also been re-enabled for select US organizations.
Source: Anthropic Blog
Big Tech Proposes Unified Jailbreak Severity Standard
The Fable 5 saga exposed a glaring gap in the AI industry: there's no agreed-upon way to measure how serious a jailbreak actually is. Now Anthropic, Amazon, Microsoft, and Google are trying to fix that.
The four companies, working through the Glasswing partnership, have proposed an industry-wide framework for scoring jailbreak severity. The system evaluates discoveries across four criteria:
- Capability gain — how much more can the model do with the jailbreak versus existing tools?
- Breadth of tasks affected — is this a narrow exploit or a wide-open door?
- Ease of weaponization — how much expertise does it take to turn the exploit into something dangerous?
- Discoverability — how likely is it that someone stumbles onto this technique independently?
The goal is straightforward: when a new jailbreak surfaces, developers, regulators, and security researchers should all be speaking the same language about how urgent it is. Without a shared standard, every new discovery triggers a scramble of conflicting assessments — exactly the kind of chaos that led to the blanket Fable 5 shutdown rather than a more targeted response.
The framework is still a work in progress and the group expects it to evolve as more partners provide feedback.
Source: Anthropic Blog
Google Bakes Computer Use Into Gemini 3.5 Flash
Google is making a significant bet on AI agents that can actually use computers. As of June 24, Gemini 3.5 Flash ships with built-in computer use capabilities — letting developers build agents that see screens, reason about what's on them, and take action across browser, mobile, and desktop environments.
This is a meaningful shift. Previously, Google offered computer use only as a standalone capability in a separate Gemini 2.5 model. Now it's baked directly into the main Flash model alongside function calling, Google Search grounding, and Maps integration. Think: continuous software testing, enterprise workflow automation, and professional application control — all from a single model.
The numbers are competitive. Gemini 3.5 Flash scores 78.4 on OSWorld-Verified, just 0.3 points behind GPT-5.5's 78.7. That's a 13.3-point jump from the previous-generation Gemini 3 Flash (65.1). And it comes at a fraction of the cost: $1.50 per million input tokens and $9 per million output tokens, compared to GPT-5.5's $5 and $30 respectively.
On the safety front, Google has implemented targeted adversarial training to mitigate prompt injection risks. Two optional enterprise safeguards let organizations require explicit user confirmation for sensitive actions and automatically halt tasks if an indirect prompt injection is detected.
Computer use is available now via the Gemini API and Google's Enterprise Agent Platform.
Source: Google Blog
EU AI Act Amendments Head for Formal Adoption
The European Union's AI regulatory landscape is shifting — and for once, the changes might actually help companies comply. The EU Parliament approved material amendments to the AI Act on June 16, and formal adoption is widely expected this month, ahead of the August 2 deadline.
The biggest practical change: high-risk AI obligations have been pushed back 16 months, from August 2, 2026, to December 2, 2027. That gives companies building AI systems in healthcare, employment, law enforcement, and other high-stakes domains significantly more time to get their compliance in order.
But it's not all relaxation. The amendments also introduce new prohibitions on "nudifier" AI applications — systems that generate or manipulate sexually explicit images, video, or audio without consent, or that create child sexual abuse material. These bans take effect December 2, 2026.
On the simplification side, the amendments clarify that AI systems used solely for user assistance, performance optimization, automation, or quality control won't automatically qualify as "safety components" — a designation that previously could trigger the full weight of high-risk compliance requirements.
The package reflects the EU's balancing act: maintaining its position as the global leader in AI regulation while acknowledging that overly aggressive timelines risk driving innovation elsewhere.
Sources: Latham & Watkins, Morgan Lewis