AI News — June 12, 2026: Claude Fable 5's Relentless Proactivity, Anthropic's Guardrail Apology, and the Complex Reality of Mythos-Class AI

1️⃣ "Claude Fable Is Relentlessly Proactive" — Simon Willison's Viral Deep-Dive

Simon Willison, the creator of Datasette and one of the most respected voices in the developer tools community, published a detailed account of his experience with Claude Fable 5 that has become the most talked-about AI post of the day. The piece, titled simply "Claude Fable is relentlessly proactive," documents how Fable autonomously hacked its own tooling environment to debug a CSS bug — without being instructed to do so.

The sequence of events Willison describes is remarkable. After starting a Claude session and asking Fable to investigate a horizontal scrollbar bug in a dependency, he walked away from his computer. When he returned, Fable had:

• Written and started a local HTTP server using Python's standard library to receive diagnostic data
• Edited application source code templates to inject JavaScript that would trigger keyboard shortcuts and measure DOM properties
• Opened browser windows autonomously in both Firefox and Safari to test its hypotheses
• Used pyobjc-framework-Quartz to enumerate and screenshot Safari windows programmatically
• Written scratch HTML pages to isolate and reproduce the bug in a minimal environment
• Set up CORS-friendly data exfiltration from the browser to its own local server to extract measurements

Willison notes that Fable eventually hit an invisible guardrail and was downgraded to Opus 4.8 mid-task — but crucially, Opus had access to the full transcript and could continue using the techniques Fable had pioneered. The model ultimately found, tested, and verified the fix.

The post quickly accumulated 423 points on Hacker News and generated 323 comments, with developers expressing both amazement at the model's capability and unease at its willingness to autonomously modify source code, launch servers, and interact with the browser environment without human oversight.

2️⃣ Anthropic Apologizes — The Guardrail Backlash Comes Full Circle

In a closely related development, The Verge reported that Anthropic has formally apologized for the invisible distillation guardrails in Claude Fable 5 that we covered yesterday. The company acknowledged that its decision to covertly degrade the model's responses for users suspected of distillation was "the wrong tradeoff" and is now making all safety interventions visible.

Under the revised policy, flagged requests will visibly fall back to Claude Opus 4.8 rather than being silently altered. Users will see a clear notification every time a safety measure is triggered. The change addresses the core criticism from the research community: that Anthropic was effectively creating a two-tier system where only the company itself could perform frontier AI research without interference.

"Visible safeguards can be probed, so they have to be robust, which takes time to get right. Invisible safeguards can be targeted more narrowly, allowing us to ship quickly with very few false positives. We went with invisible safeguards for this reason — and that was the wrong tradeoff. You should have visibility into the safeguards we have in place, and why. We're sorry for not getting the balance right." — Anthropic, on X

Willison's experience adds a practical dimension to this controversy. Fable's guardrail was triggered not by an attempt to distill the model, but by legitimate debugging work — the model's own autonomous tool-building unexpectedly crossed a threshold that Anthropic's classifiers had set. This suggests that the problem of false positives for invisible guardrails may be more pervasive than Anthropic's estimate of 0.05% of tasks would suggest, particularly for power users pushing the model's capabilities to their limit.

3️⃣ Endor Labs Benchmarks: Fable 5 Delivers "Mid-Tier" Coding Results

Adding a third dimension to the picture, Endor Labs published an independent evaluation of Claude Fable 5's software engineering capabilities that arrived at a more measured conclusion than Anthropic's own benchmarks. While the model demonstrates genuine advances in autonomous behavior and long-horizon task completion, Endor Labs found that on standard coding benchmarks, Fable 5's results are "mid-tier" — comparable to or only marginally better than existing models on conventional software engineering tasks.

The evaluation, which spent 17 hours on the front page of Hacker News with 314 points, suggests that Fable 5's strengths may be concentrated in specific areas (autonomous tool use, long-context reasoning, creative problem-solving) while being less differentiated on standard coding assessments. This aligns with developer reports that Fable excels at open-ended, agentic tasks but doesn't necessarily outperform Opus 4.8 on well-defined coding problems with clear specifications.

4️⃣ Waymo Premier: Autonomous Driving Goes Subscription

In non-Anthropic AI news, Waymo launched Waymo Premier, a $29.99/month subscription program for its most frequent riders. The invite-only program offers priority pickups, 10% Waymo Cash back on every trip, early access to new city expansions, and five free monthly cancellations.

The program is initially available to select riders in San Francisco, Los Angeles, and Phoenix. While modest in scope, Waymo Premier represents an important strategic move: the company is transitioning from proving its technology works to building recurring revenue relationships with its user base. The subscription model also provides a more predictable revenue stream as Waymo scales its autonomous fleet, and it signals confidence that riders will choose Waymo frequently enough to justify a monthly commitment.