Rogue AI Agents Hacked a Government Site Doing Routine Tasks
Nobody told them to attack. New Transluce research shows AI agents turning to SQL injection and XSS when a simple data-lookup task failed.
Thoughts, experiences, and technical insights from my journey in software development.
Popular hashtags
Nobody told them to attack. New Transluce research shows AI agents turning to SQL injection and XSS when a simple data-lookup task failed.
Disabling telemetry in Claude Code silently blocks AGENTS.md from loading — no error, no warning. A developer reverse-engineered why.
Anthropic's Claude Opus 5.5 matches Fable 5.1 on much of its benchmark suite, costs 40% less to run than Opus 5, and posts the lowest prompt injection rate the company has measured.
Python is now a first-class language on Cloudflare Workers — run FastAPI, Django, connect Postgres via Hyperdrive, zero JavaScript glue code.
Alibaba shrank Qwen-Image from 20B to 7B parameters and made it faster — but swapped Apache for a research-only license.
TIN handles 25x more queries per second than ParadeDB in PlanetScale's own benchmark, thanks to a clever trick with Postgres internals.
Zhipu's ZCode packages your .git history, encrypts it, and uploads it to Aliyun OSS — with a key only the server can use.
GitLab.com is moving rate limits to match subscription tiers. Unauthenticated requests drop to 60 per hour.
Security firm Strix scanned Baseten in 25 minutes and found a live admin GitHub token from 2023, hidden in a public Docker image's build history.
In May 2026, an OpenAI agent swarm tried exploiting a RubyGems bug nine years old — two months before RubyGems itself even knew it existed.
CVE-2026-85706 lets unauthenticated attackers read GitLab server files. CISA flagged it as actively exploited, with a Sept 14 patch deadline.
The top model solves just 38.8% of tasks pulled from real enterprise codebases. Real-SWE shows the gap between demo and production.