How Ultrafast cut incident investigation from hours to minutes at OpenAI
- ID
- 39667
- Status
- summarized
- Published
- 26 Oct 2026, 3:00 PM
- Fetched
- 27 Oct 2026, 3:54 AM
- Provider
- OpenAI News
- Category
- ai-labs
- Original URL
- https://openai.com/index/openai-cut-incident-investigation
- Source URL
- https://openai.com/news/rss.xml
Summary
- Score
- 4.0
- Created
- 27 Oct 2026, 3:54 AM
- Tags
- Audience
- developersai_ml_learnersai_agent_users
What happened
OpenAI published an internal-usage post describing how its own engineers use "Ultrafast," a high-speed mode in Codex, as the default for incident response. AJ Trbojevic, a Member of Technical Staff and Datacenter Security Architect, says security investigations that used to take one to two hours to pull structured and unstructured data together now take 10 to 15 minutes, sometimes near real time; Aaron Friel describes pointing Codex at a fresh incident Slack channel to monitor and investigate multiple possible causes at once. No benchmarks, cost figures, or independent verification are included, and the post states it is part of a series of illustrative internal examples.
Why it matters
Treat the 1-2 hours to 10-15 minutes figure as an unverified vendor claim, but the reusable pattern is concrete: an agent wired into the incident Slack channel testing several failure hypotheses in parallel instead of one engineer serialising them. If you want to copy it, the work is not the model — it is granting an agent read access to logs and Slack history during an incident, which is a data-exposure and permissions decision you have to make before the next outage, not during it.
Discussion angle
Would you let an agent sit in your incident channel and read production logs unattended? Walk through what access you'd grant, what it would have to log, and whether 10-15 minutes versus 1-2 hours changes your on-call rotation or headcount math at all.