Anthropic accidentally exposes 512,000 lines of Claude's internal source code via an npm package map file, revealing unreleased features, secret codenames, and a more powerful Claude 4.6 variant — marking the company's second major accidental disclosure in two weeks and raising serious concerns about the safety-focused AI firm's own operational …
A new survey reveals 91% of software engineers now use AI coding tools daily, with 42% saying they'd hesitate to hire developers who don't, as AI proficiency rapidly becomes a non-negotiable career skill rather than a threat to developer jobs.
Anthropic accidentally leaks 512,000 lines of Claude's proprietary TypeScript source code via an npm package, exposing secret AI architecture, unreleased model codenames, a hidden 'Undercover Mode,' and an internal model with a 29-30% false claims rate — while users who installed the package during a specific window face additional supply-chain …
AI recruiting startup Mercor confirms it was hit in a massive supply chain attack on open-source library LiteLLM, which sees millions of daily downloads, with extortion group Lapsus$ claiming responsibility and sharing allegedly stolen data including Slack conversations and contractor videos, while thousands of other companies may also be compromised.
Google DeepMind releases two new tools — the Gemini API Docs MCP and Gemini API Developer Skills — that together achieve a 96.3% pass rate and 63% fewer tokens per correct answer, keeping AI coding agents current with real-time Gemini API documentation.
Slack launches 30+ AI features, including cross-platform meeting transcription for Zoom and Google Meet, an expanded Slackbot with multi-step task automation, MCP enterprise tool integration, and a native CRM, positioning itself as an all-in-one AI productivity hub available to all Free and Pro users in April.
Agent harnesses — the software infrastructure surrounding AI models that manages memory, tools, and planning — are quietly becoming the true determinant of real-world AI performance, with implementations like Anthropic's Claude Agent SDK proving that a well-designed harness can outperform raw model capability alone.
A Northwestern University study reveals over 60% of federal judges now use AI tools in their judicial work, yet nearly half report receiving no AI training and policies governing its use remain dangerously inconsistent across courts.
Hermes Agent v0.6.0 launches March 30, 2026, bringing multi-profile isolation, expanded platform integrations including Feishu, WeCom, and Slack multi-workspace OAuth, plus a major security overhaul with atomic config writes, dangerous command detection, and automatic inference failover across 95 merged PRs.
DeepSeek suffers its longest outage in over a year, with two incidents totaling more than eight hours of downtime on March 29-30, 2026, just as the Chinese AI startup prepares to launch its next major model, V4, a multimodal system trained with help from Huawei and Cambricon.