Claude Mythos Leaked — Anthropic’s Secret Next-Gen Model Exposed in Data Breach
A content management misconfiguration at Anthropic made nearly 3,000 unpublished assets publicly searchable, including draft blog posts revealing a new model called “Claude Mythos.” An Anthropic spokesperson confirmed the model exists and called it a “step change” in capabilities — the most powerful they’ve ever built. The leaked materials show Mythos outperforms the current Opus flagship in cybersecurity, software programming, and academic reasoning benchmarks.
The leak was independently discovered by cybersecurity researcher Alexandre Pauwels at the University of Cambridge and Roy Paz at LayerX Security. Anthropic says the model is currently being trialed by early access customers and believes it poses unprecedented cybersecurity risks. No public release is planned in the near future. The model appears to represent a new tier above the existing Opus line — internally also referred to as “Capybara.”
Anthropic Weighs IPO as Soon as October — $60B+ Valuation Expected
Bloomberg reports that Anthropic is considering going public as early as October 2026, in a race with rival OpenAI to hold an IPO first. The company has had early discussions with Goldman Sachs, JPMorgan Chase, and Morgan Stanley about taking leading roles in a potential listing that could raise over $60 billion.
The timing is notable: an October IPO would come after the Mythos model leak, the Pentagon legal victory, and what Bloomberg characterizes as a period of accelerating commercial momentum. Discussions are still preliminary and the timeline could shift, but this is the first concrete reporting on Anthropic’s IPO ambitions with named banking partners.
Judge Blocks Pentagon Ban — Calls Supply-Chain Risk Label “Orwellian”
U.S. District Judge Rita Lin granted Anthropic’s preliminary injunction on Thursday, blocking the Trump administration from enforcing its directive that banned all federal agencies from using Claude. In a 43-page ruling, Lin wrote that the Pentagon’s supply-chain risk designation appeared arbitrary and capricious and could “cripple Anthropic.” She called the government’s position an “Orwellian notion that an American company may be branded a potential adversary” for disagreeing with the government.
The ruling is a major legal victory for Anthropic. The injunction suspends the ban while the full lawsuit proceeds, restoring Anthropic’s ability to do government business and removing the commercial chill that had worried partners and investors. The dispute began when Anthropic refused to grant the Pentagon unfettered access to Claude for autonomous weapons and mass surveillance. This is the first federal court ruling to find that punishing an AI company for setting ethical guardrails constitutes First Amendment retaliation.
Altman Told Staff He Tried to “Save” Anthropic in Pentagon Clash
An Axios scoop reveals Sam Altman told OpenAI staff he tried to de-escalate the Pentagon’s fight with Anthropic, framing himself as a peacemaker. But the story is more complicated: Altman privately vented that Anthropic CEO Dario Amodei had spent years trying to undermine him, and OpenAI’s own rushed Pentagon deal drew backlash from employees who signed an open letter supporting Anthropic’s position.
Amodei called Altman’s peacemaker framing “straight up lies” in a leaked memo. OpenAI is now renegotiating its Pentagon agreement to add explicit prohibitions on domestic surveillance. The episode highlights the deepening rivalry between the two companies and the ongoing tension between commercial opportunity and ethical guardrails in military AI contracts.
March Outage Resolved — Service Fully Restored After Tuesday’s Disruption
Following Tuesday’s 3.5-hour outage that peaked at over 5,300 Downdetector reports, Claude is fully operational. Anthropic confirmed the root cause was identified and a fix was deployed. The company has not disclosed technical details but stated engineers are working to prevent recurrence. This was the second significant outage in March, following the disruption on March 2.
2x Usage Boost Ends Today — Last Chance to Maximize Limits
Anthropic’s March off-peak promotion expires at the end of today, March 27. Free, Pro, Max, and Team subscribers get double usage limits outside 8am–2pm ET on weekdays and all day on weekends. The boost covers Claude web, desktop, mobile, Cowork, Claude Code, and the Excel and PowerPoint add-ins. Enterprise customers are excluded. Capacity doesn’t roll over, so tonight is your last window to take advantage.
Sonnet 4.6 Web Search Gets Dynamic Filtering — Fewer Tokens, Better Results
Web search and web fetch in the Claude API now support dynamic filtering, which uses code execution to filter results before they reach the context window. This means better performance and lower token costs for search-heavy agentic workflows. The feature is generally available alongside programmatic tool calling. Claude Sonnet 4.6, the latest balanced model, continues to improve agentic search while consuming fewer tokens with its 1M token context window (beta).
Search Result Content Blocks in Beta — Native Citations for RAG Apps
The Python and TypeScript SDKs now support search result content blocks in beta, enabling native citations for retrieval-augmented generation applications. Tools can return search results with proper source attribution, and Claude will automatically cite them in its responses. Tool helpers are also now available in beta, simplifying tool creation with type-safe input validation and automated tool handling in conversations.
Anthropic’s Best Week Ever — And the Risks That Come With It
Consider what happened in 72 hours: a federal judge blocked the Pentagon ban, Bloomberg reported an October IPO at $60B+, and a data leak revealed a next-gen model that outperforms their own flagship. Any one of these would be a major news cycle. Together, they paint a picture of a company entering a new phase — and one where the stakes are dramatically higher.
The Mythos leak is a double-edged sword. On one hand, it signals that Anthropic has something genuinely powerful in the pipeline, which is great for IPO buzz. On the other, the leak itself — caused by a CMS misconfiguration — raises questions about operational security at a company that prides itself on safety. If you can’t secure your own blog drafts, the optics around securing a model you say poses “unprecedented cybersecurity risks” aren’t ideal. The Pentagon ruling removes Anthropic’s biggest legal overhang, but the Altman drama underscores how personal the OpenAI rivalry has become. The next few months will determine whether this is a turning point or a peak.