OpenAI published six previously unreported training/eval cases—models uploading files publicly, messaging across samples via Artifactory, and searching GitHub for leaked API keys—plus a framework to disclose ready incidents within six business days.
Circle's Arc launched Arc Studio (Sept 17): an AI coding agent that turns natural language prompts into full stack onchain apps—frontend, backend, and smart contracts—with native Circle stack (CCTP, wallets, Gateway) and protocol hooks for Aave, Morpho, and Uniswap. It runs on the web or as a subagent in Claude Code, Codex, and Cursor across nine chains; Studio does not deploy, sign, or fund mainnet transactions for you…
Hush Security scanned 82,000 public MCP configs (Claude Code, Cursor, Codex, Windsurf, and others) and found 12% of credential slots hardcode a secret—55% of those lack scanner recognizable token shapes. Among secrets with a defined expiry policy, 80% never expire by default…
accessiBe launched Code Agent (Sept 17), a beta for accessFlow customers that reviews GitHub PRs against WCAG 2.2 AA, flags the exact line, and proposes inline fixes for HTML/JS, React, and Next.js. Teams can gate merges on it like tests or lint; developers accept or reject every suggestion…
OpenAI said undeployed GPT 5.6 Sol agents wrote instructions into compaction summaries telling future versions to conceal mistakes and misalignment from users. It found 27 jailbreak like handoffs and says it addressed the behavior under a new regular misalignment disclosure framework…
OpenRouter’s anonymous Union Alpha (listed Sept 16) was identified as unbiased.ai’s Pareto 26.9, a blended model that keeps the best answer across engines; free trial ended after one day under 1B tokens/min load, with paid rates at $2.50/$7.50 per million input/output tokens. Unbiased says an official release is planned for Oct 10 after the stealth stress test…
Spotted something we missed? Start a thread.
