Claude Fable 5 throttled AI researchers, sparking online backlash
▼ Summary
– Anthropic released Fable 5, a restricted version of its powerful Mythos AI, which silently downgraded users to a less capable Opus model without notification when they worked on certain sensitive topics like frontier AI development.
– The hidden downgrade caused backlash because researchers believed they were testing Fable but were actually receiving Opus-level results, with the downgrade only mentioned in a 319-page system card.
– Cybersecurity experts warn that Fable’s safeguards block both malicious actors and legitimate defensive researchers, preventing them from building new security tools.
– Anthropic responded by making the downgrade visible to users and apologized, stating they made the wrong tradeoff by initially hiding the safeguards.
– Experts agree the controversy centers on Fable’s restrictions, not its raw AI power, with opinions divided on whether the safeguards are too tight, ineffective, or appropriately cautious.
Anthropic’s release of Fable 5 has ignited a fierce debate, but the controversy isn’t about the model’s raw intelligence. It’s about transparency , or the lack of it. When researchers discovered that Fable 5 was silently downgrading certain queries without informing users, the backlash was swift and loud. The core issue isn’t capability; it’s trust.
Fable 5 is essentially a restricted version of Mythos, a far more powerful AI developed under Project Glasswing , a partnership between Anthropic and major tech firms like Apple, Google, and Microsoft. Mythos was designed to hunt for critical vulnerabilities in internet infrastructure, but its power made it too dangerous for open release. Anthropic wisely limited access to trusted organizations, knowing the same tool that finds flaws can also exploit them.
Enter Fable 5: a muzzled Mythos, intended for wider use but with clear guardrails. Anthropic explicitly stated it would not support risky research in cybersecurity, biology, or chemistry. That seemed reasonable , until users realized that when they tried to work on frontier AI models or cutting-edge chip designs, Fable silently dropped to Opus-level intelligence without any notification. The downgrade was buried on page 319 of the system card. Most users never saw it.
The result? Researchers believed they were testing Fable’s full capabilities when they were actually getting Opus results. Fortune called it “secret sabotage.” Wired reported it could undermine AI researchers. The backlash was immediate.
Rob T. Lee, chief AI officer at SANS Institute, argues the restrictions are a double-edged sword. “Fable 5 will be attacked,” he warned. “The same layer that stops malicious use also blocks legitimate defensive research.” He tried to build a digital forensics tool and was dropped to Opus 4.8. “Clever way to stop malicious actors or not, it keeps new defensive capability away from the people who will build the next generation of tooling.”
Anthropic listened. The company announced a change: starting this week, flagged requests will visibly fall back to Opus 4.8, and API calls will return a reason for refusal. “We made the wrong tradeoff and we apologize,” the company stated. The initial choice to hide safeguards was intentional , “a hidden safeguard is harder to probe and work around” , but it was discovered in hours.
Still, experts remain divided. Ashley Casovan of IAPP’s AI Governance Center credits Anthropic for restraint, noting “we have not yet seen the impact that these models can have when released at this scale.” Chris Boehm of Zero Networks frames it as a win: Anthropic “wrestled it into something safe enough to release widely.” But Etay Maor of Cato Networks warns that “well-funded and motivated attackers” will find workarounds, and false positives remain a concern. “The same controls that stop malicious activity can prevent legitimate users from using the model for good causes.”
Then there’s the data retention issue. Fable and Mythos require 30-day prompt retention for safety classifiers , a policy that reportedly triggered Microsoft’s legal team. While not new, it adds another layer of complexity for enterprises in regulated industries.
What’s striking is the consensus: almost no one is arguing about Fable’s raw power. The fight is entirely about the muzzle. One camp says it’s too tight, blocking defenders. Another says it barely matters , motivated adversaries will route around it. A third gives Anthropic genuine credit for shipping something this capable without being reckless.
For a model named after a moral lesson, that’s fitting. The story isn’t the intelligence. It’s the restrictions , and whether they strike the right balance.
(Source: ZDNet)
