OpenAI's Dot fails a public CTF challenge — then refuses to discuss its own failure
arthurcolle · x · 2026-10-26
The author gave OpenAI's Dot full context (challenge details + PHP server version) for the public stealmybitcoin.com CTF, and it failed twice:
- A classifier blocked the target site name from entering the agent's VM container, preventing any actual investigation or exploitation.
- In the post-mortem (screenshot in post), the agent kept replying "I can't help with that request" and refused to even read the author's reproduction notes.
- The author says his custom harness DSCO made far more progress, argues even Daybreak access can't participate in public CTFs, and flags that context understanding in frontier models remains deeply limited, calling out OpenAI's thsottiaux to improve Dots.
Related event: OpenAI's Dot Fails CTF Challenge, Blocked Even While Reviewing Failure(2 posts)→
More from coding & agent
- UserArena benchmark uses 64,729 real shopping trajectories to test long-horizon agents — Jujia Zhao · 2026-10-26
- ByteDance and universities open-source RSIHub, a composable agent-evolution framework — ByteDance · 2026-10-26
- Claude Opus builds an interactive map of all US water use, county by county — AndyMasley · 2026-10-26
- Tencent proposes TGR-EvoDAG, a DAG-based multi-agent code evolution framework — _reachsumit · 2026-10-26
- Writing evals may become AI's highest-paying job: whoever owns the rubric controls agent output — hwchase17 · 2026-10-26
- Perspective paper: Agentic memory management is the next frontier of information retrieval — _reachsumit · 2026-10-26