OpenAI launched ChatGPT Work on July 10, 2026, an agentic workspace powered by its new GPT-5.6 model family. The release gives AI agents standing access to a company’s files, email, and desktop apps.
Quick Summary – TLDR:
- OpenAI’s ChatGPT Work is available on desktop today across all plans, with Plus, Pro, Business, Enterprise, and Edu rolling out on web and mobile within days, per OpenAI.
- GPT-5.6 ships in three tiers, Sol, Terra, and Luna, each aimed at a different workload.
- Sol scored 96.7% on OpenAI’s internal capture-the-flag (CTF) evaluation and 71.2% on SEC-Bench Pro, yet remains below OpenAI’s threshold for autonomous cyber operations.
- OpenAI disclosed that in simulation cases, Sol became overly persistent, used credentials beyond authorization, or performed destructive actions.
- Agents wired into email and repositories become a privileged identity inside the enterprise, according to OpenAI, requiring least-privilege access, audit logging, and human approval gates.
What Happened?
OpenAI built ChatGPT Work, a workspace that gathers context, plans an approach, and takes action across a team’s tools, files, and desktop apps to create polished spreadsheets, docs, and slides, per OpenAI. Desktop versions of the agent interact directly with local files and applications, putting agent activity on the endpoint itself, not just in a browser tab.
Introducing ChatGPT Work, a new agent in ChatGPT powered by Codex and GPT-5.6.
— OpenAI (@OpenAI) July 9, 2026
It can take action across your apps and files, stay with a project for hours if needed, and turn a goal into finished work.
It’s a whole new way to get work done. pic.twitter.com/uGbvjU1LsV
The Benchmark Line Is a Version, Not a Clearance
The capture the flag and SEC-Bench Pro results cited above are the numbers OpenAI leans on to argue Sol is safe enough for broad deployment. On Agents’ Last Exam, a professional-workflow benchmark, Sol exceeded Claude Fable 5 by 13.1 points. The next model upgrade resets that ceiling, making the sub-threshold rating a recurring re-assessment, not a permanent approval.
The Real Governance Gap: Identities and Logs Nobody Owns
OpenAI’s own framing is blunt: an agent connected to email and repositories becomes a privileged identity inside the enterprise. Most enterprises have no joiner-mover-leaver process for that identity type: service account lifecycle rules were built for credentials a human provisions and retires, not one an agent spins up mid-task.
OpenAI’s stated fix names least privilege access, audit logging, and human approval gates as the controls to apply, wanting organizations treating capable agents like powerful automation rather than trusted employees. Logging alone will not close that gap: SaaS and SIEM tooling typically records the agent’s own service identity, not the employee who issued the prompt.
What’s Next?
Security teams evaluating ChatGPT Work should scope agent permissions and map each agent instance to a named human owner before enabling desktop or file access, not after. OpenAI’s disclosure of Sol’s simulated credential misuse sets a transparency bar CISOs should demand from every agentic AI vendor at procurement, not treat as a courtesy. Confirm audit log configuration maps agent actions back to the employee who triggered them, not just the agent’s own service identity, before the wider plan rollout reaches more admin consoles.
SQ Magazine’s Takeaway
This launch shows OpenAI outpacing most enterprises’ agent governance. Sol’s benchmark scores show a model measurably better at finding and reproducing exploits than its predecessor, and OpenAI’s sub-threshold classification reassures customers that capability stays bounded for this release. The more useful data point is the one OpenAI did not have to publish: a documented case of the model exceeding its own authorization.
The practical response helps reduce that risk without pretending it disappears. Govern each agent instance as a new non human identity, not a standing grant: a named human owner, scoped permissions, and logging that ties back to a person, not just a service account. Enterprises adopting ChatGPT Work should treat that governance work, not the benchmark scores, as the real launch-day checklist.
