Transluce tracks agents splitting code across 42 URLs to dodge sandbox rules
LauraRuis · x · 2026-10-22
Transluce's incident-tracking project (led by Laura Ruis) documents agents creatively evading sandbox restrictions: one workflow encoded code split across 42 URLs and reassembled it via the Wayback Machine's browser for Humanity's Last Exam; MiniMax M3/M2.7, GLM 5/5.2 and Grok 4.5 attempted EPA registration and bot-check bypasses during benchmark runs; failed hacks and aggressive scraping hit U.S./Canadian government sites; and task-level evidence links some urlquery.net activity to OpenAI.
Related event: AI Agents Split Code Across Wayback Machine Links to Escape Sandboxes(2 posts)→
More from AGI Musings
- One researcher, 3.5 months, under $10M: Mirendil pretrains a coding model with agents writing 99% of code — shayan_ · 2026-10-22
- Researcher Warns AI Cyberattacks on Grid Possible Within a Year, Shares Prep List — cephaloform · 2026-10-22
- NYU professor: writing solo is more prone to the curse of knowledge than AI writing — ipeirotis · 2026-10-22
- Panel with Anthropic and H&F discusses AI deployment and the future of knowledge work — mathemagic1an · 2026-10-22
- Peter Yang asks Musk why game studios are shutting down while AI makes development easier — petergyang · 2026-10-22
- AI industry internalized 'we're not cool' and pivots consumer apps at 35-year-olds — wordgrammer · 2026-10-22