Issue #2 | DailyPrompt

01 research

UK Government Confirms Claude Mythos Is the First AI to Complete an Autonomous Cyberattack End-to-End

The UK's AI Security Institute published an independent evaluation of Claude Mythos Preview, finding it is the first AI model ever to complete a full 32-step network takeover simulation from start to finish — succeeding in 3 of 10 attempts and averaging 22 steps. On expert-level cybersecurity tasks that no AI could solve before 2025, Mythos scored 73%, a dramatic leap over any previous model. The findings, from a government body rather than Anthropic itself, lend independent weight to the company's case for keeping Mythos restricted to vetted security partners.

AISI →

product

Anthropic Quietly Launches Passport and Selfie Checks for Certain Claude Features

Anthropic has rolled out identity verification for users seeking access to specific Claude capabilities, requiring a government-issued passport scan and a selfie submitted through identity provider Persona. The company says the checks are designed to prevent abuse, enforce usage policies, and meet legal compliance obligations. The launch drew attention partly because it was not announced publicly — users discovered it only when prompted mid-session.

Claude Help Center

Anthropic

@AnthropicAI

Research we co-authored on subliminal learning—how LLMs can pass on traits like preferences or misalignment through hidden signals in data—was published today in @Nature.

Apr 15 View on X →

community

$ cat story_03.md

OpenClaw Creator's Claude Account Was Suspended by Anthropic — Then Restored Within Hours After Going Viral

Peter Steinberger, creator of the popular third-party Claude Code client OpenClaw, posted on X that Anthropic had suspended his personal Claude account for "suspicious activity" — an action that came days after Anthropic announced subscriptions would no longer cover third-party tools. The post spread rapidly, drawing criticism from the developer community, and Anthropic reversed the suspension within hours. An Anthropic engineer reached out directly to clarify that the company had never banned anyone specifically for using OpenClaw.

$ open techcrunch →

announcement No. 04

Europe Backs Anthropic's Decision to Keep Its Most Powerful Model Away from the Public

The European Commission publicly endorsed Anthropic's choice to withhold Claude Mythos from general release, citing the model's demonstrated ability to independently discover and weaponise zero-day vulnerabilities across every major operating system and browser. Anthropic says it chose a restricted rollout through Project Glasswing precisely because of these capabilities, and the EU's backing marks one of the first times a major regulatory body has explicitly supported an AI company's decision to limit its own model's availability on safety grounds.

Source: NBC News →

community

Anthropic Accidentally Wiped Thousands of Innocent GitHub Repos While Chasing a Leak

In an attempt to scrub its own leaked source code from the internet, Anthropic issued takedown requests that resulted in thousands of unrelated GitHub repositories being removed — a move the company subsequently described as an accident. The incident exposed the blunt-instrument nature of automated DMCA enforcement on major code-sharing platforms, and left developers scrambling to recover repositories that had no connection to the leaked material.

TechCrunch →