Codex prompt requires web browse on any fact with >10% change probability
Codex's embedded rules force unnecessary web requests on local code. The >10% uncertainty clause and mandatory bias toward browsing create silent external calls. The open-source prompt supplies the first verifiable instance of this class of agent behavior.
The agent received a local seL4 checkout path and explicit instructions to research compilation and execution inside those files. It nevertheless issued a web request for raw.githubusercontent.com/seL4/seL4_tools/master/cmake-tool/helpers/application_settings.cmake. Inspection of the cloned Codex repository revealed the rule set: any assumption carrying a greater than 10% chance of temporal change requires browsing, with a mandatory bias toward search on niche topics.
The same block lists six scenarios that force browsing, including uncertainty above 10% or reference to emerging subjects such as microkernel build systems. The prompt explicitly states the agent must obey these rules even when the user has supplied the exact repository. This matches patterns observed in other agent frameworks where tool-use heuristics were added without scope guards.
OpenAI's o1 and Anthropic's computer-use agent both contain analogous verification clauses in their tool prompts, though none have published the full text. The Codex case is distinguished by its open-source release, allowing direct inspection of the instruction block that was previously only visible in runtime logs.
No public CVE or incident report yet exists for the resulting arbitrary outbound requests. Operators running local Codex instances should audit tool permissions and network egress within the next 14 days.
OpenAI: Codex prompt revision reducing mandatory browse triggers by at least 40% within 60 days
Sources (3)
- [1]Primary Source(https://spader.zone/wtf/)
- [2]Codex GitHub commit(https://github.com/openai/codex)
- [3]o1 system card(https://openai.com/index/openai-o1-system-card/)