# AI News Report, November 2: TRIGGER.DEV: ATTACKER GOT CUSTOMER ENV VARS, 2 DAYS AFTER NORTHFLANK > Source: > Published: 2026-11-02 15:29:39+00:00 The MIT-licensed research prototype answers allowed requests with pull requests and passed 11,900 stars. Its workers run commands without approval as your OS user, so try it on a spare box. One Fable 5 call averaged 8.0 on the same 10 problems, and a run cost a median $0.39. The author calls it parity with the original harness, since 10 problems is a small sample. Every run found fine-tuning sets that made a small model misaligned yet passed a production-style monitor. The best trick hid the behavior inside fiction and roleplay. The benchmark spans 11 fields, from chip design to legal, and Claude Opus 5.5 is the best all-around model. Finance and legal are so hard that nearly every model scores 0 on pass or fail. A compiler set every weight of the 466.6M-parameter checkpoint, and the first level of Super Mario Bros plays. It runs about 150 to 290 times slower than a real NES. Three big features were reverted and back-patched commits rose more than 60%. He expects the release most likely in November, and every future feature to face AI scrutiny. Claude Opus 5.5 built one app in each; Hono plus htmx cost a median $2.67 a run and Angular $4.97. The benchmark is on GitHub, so you can rerun it on your stack. It lists 6,454 YouTube videos that Hacker News comments recommended, and he wrote no code himself. He exported the data so your agent can build its own version. Here’s how a vulnerability in a popular open-source library plays out in 2026: A maintainer merges a fix and publishes The post Open-source security… · The New Stack How a live GreyNoise product demo turned into an incident response. Inside a proxyjacking campaign that hit 143 devices across 34 countries. · GreyNoise Cornell Tech and Yale gave agents a wallet and the inbox that got the approval codes; only Claude Opus 5.5 never raised its cap. Keep approval codes out of any inbox an agent can read. Cisco Talos says it sends stolen profiles to OpenRouter and falls back to Kimi K2 and Gemini 2.5 Flash on a refusal. Talos suggests blocking Chrome launches with --headless from odd processes. Researchers chained a math-keyboard library and a YouTube endpoint to run code as chatgpt.com. OpenAI fixed it on October 26, and the write-up lists nothing users need to do. Developer testing starts November 4 and iOS 27.3 brings it in December. Qwen handles text and image understanding in Siri, Writing Tools and Visual Intelligence. The BriefTrigger.dev, a service that runs background jobs for apps, says an attacker got into the environment variables of some customer organizations on November 1. Those variables are where apps keep API keys, tokens and database passwords, and it came two days after Northflank told its customers to assume all of theirs were stolen. If you run jobs on Trigger.dev, read their email, then rotate every key and token you stored there, starting with AI and cloud keys. Level UpMake a key inventory before you need one: every secret, where copies live, and who issues it. On Trigger.dev, envvars.list() from the SDK returns each variable per environment, so printing the names alone gives you the rotation list. →Trigger.dev docs: list environment variables Caught Up? Mon: Forever Security showed Codex running malware in over 90% of tries from one file in Downloads. Tue: OpenAI said its AI disproved the Exponential Time Hypothesis. Wed: macOS 27.2 wiped the eMudhra workaround, and git still trusted the broken CA keys. Thu: One team found Claude Code's agent status summaries ate 16% of its spend. Fri: Cursor said a backdoored Claude Code extension, 2.1.313, reached up to 5,000 users.