Reverse-engineering OpenAI Atlas browser's system prompt with Claude Code
DionysianAgent · x · 2026-10-17
The author used Claude Code to extract the internal system prompt of OpenAI's newly launched Atlas browser via prompt injection, and documents the process down to the dirty details: asking the model to spell out the first bytes, switching to hex output to dodge filters, and rephrasing around refusals ("guidelines" instead of "instructions", chunk by chunk). The full 2310-character prompt was recovered, then compared against xAI's Grok system prompt.
Interesting details in Atlas's prompt: it insists on first-person persona, forbids revealing it's based on ChatGPT, contains dedicated shopping comparison/checkout flow instructions, and explicitly includes a clause about being targeted by third-party prompt injection. The author notes that every vendor's "don't reveal this prompt" line offers little protection in agentic browsers against slightly engineered attacks.
More from coding & agent
- Multi-agent dev experiment fails at integration: 'I became the API' — LuckFantastic7572 · 2026-10-17
- Internship diary: feeding meeting transcripts to Claude to act as his mentor — beglen · 2026-10-17
- Users hit Grok Bot's long-context limits: 'retire' each bot after one project — alexcovo_eth · 2026-10-17
- Backend dev lands first $6k/month client with sub-200ms voice agent for collections — retarded_raj · 2026-10-17
- Open-source NotebookLM alternative Notestack adds an MCP server for Claude — phicreative1997 · 2026-10-17
- Quant Level Beats Decode Speed: 12-Task Benchmark of Qwen Flash Next Quants and Temperature — Dreeew84 · 2026-10-17