Anthropic: Model leaks, safety holds, API regressions, ethics crises
Key Questions
What is Anthropic's Claude Mythos model?
Claude Mythos is an advanced AI model that tops benchmarks and has identified thousands of zero-day vulnerabilities in major operating systems and web browsers. It sparked a race to fix critical bugs, some unpatched for decades. Leaks revealed its power, prompting safety concerns.
Why did Anthropic limit the rollout of Mythos?
Anthropic restricted access due to fears that hackers could use the model for cyberattacks, enabling faster attacks than ever. They want companies to implement defenses first. The model is described as 'scarily powerful' and unreleased publicly.
What caused the Claude.ai outage on April 13?
Claude.ai experienced a downtime incident, with the team posting an update that they are investigating the issue. It was highlighted on Hacker News with 16 points. No further resolution details were provided in the reports.
What happened to Claude Opus 4.6's performance on BridgeBench?
Accuracy on the BridgeBench hallucination test dropped from 83% to 68%. This regression was discussed on Hacker News with 63 points. It indicates increased hallucinations in the model.
Why was Anthropic's cache TTL downgraded?
Around March 6, cache TTL was silently reduced from 1 hour to 5 minutes, leading to quota inflation and 17-25% higher costs. It was reported on Hacker News with 232 points. Users noted significant impacts on API usage.
What API regressions have affected Claude?
Claude experienced nerfs, throttling, and a shift in user sentiment, with discussions on Hacker News. OpenClaw was banned in subscriptions, boosting GPT-5.4 usage by 8.9%. Tools like a 'digital whip' emerged to improve speed.
What regulatory actions are targeting Anthropic?
Anthropic faces a Treasury summons and a $5B Pentagon suit. These ethics crises were covered in Stratechery. They coincide with safety holds on models like Mythos.
What is Project Glasswing?
Project Glasswing involves holds or locks prompted by zero-day vulnerabilities found by models like Mythos. It relates to Anthropic's safety measures against AI-enabled cyber threats. Details emerged amid model leaks and benchmarks.
Mythos tops benchmarks but zero-days prompt holds/Glasswing locks; cache TTL slashed Mar6 (17-25% costs); Claude nerfs/throttling; Opus 4.6 hallucination drop 83%->68% BridgeBench; Apr13 outage; Treasury summons, $5B Pentagon suit; Stratechery.