If your team is using AI for real work, the risk isn't only a bad answer. It's sensitive context, internal judgment, or process know-how leaking into places you didn't intend.
That usually doesn't happen through one dramatic breach. It happens through routine use: too much copied into chats, too many people sharing one workspace, and too few rules around what AI can reuse or reveal.
OpenAI tightens defenses against adversarial distillation
OpenAI outlined new defenses against adversarial distillation, the practice of using a model's outputs to train another model without equivalent safeguards. The company said newer protections in GPT-6.2 Astra and GPT-6.2 Sol focus especially on preventing extraction of protected reasoning traces, which it argues can transfer advanced capabilities more effectively than summaries alone.
It also described a two-level approach: blocking risky requests and outputs inside the API, then looking for coordinated campaigns across accounts. OpenAI said it is also applying selected improvements back to GPT-6 Astra.
- Reasoning extraction blocked more often. OpenAI said internal adaptive testing on GPT-6.2 Astra cut average attack success from 54.0% on GPT-6 Astra with original safeguards to 0.2% with updated safeguards.
- Protections now cover full interactions. The company said its safeguards check both assistant messages and tool-call arguments, because protected reasoning can leak through tool use as well as final answers.
- Reuse limits got stricter. OpenAI said GPT-6.2 Astra and Sol tie protected reasoning to the conversation history that produced it, and drop later reasoning if earlier messages are rewritten.
- Campaign detection expanded. Beyond single requests, OpenAI said it investigates coordinated extraction attempts across accounts and has been taking weekly enforcement actions at scales from thousands to tens of thousands of accounts.
The business issue is containment
For a business owner, the headline isn't model theft. It's a reminder that AI systems can expose more than final answers. They can also expose patterns of judgment, internal methods, and operational context that are valuable precisely because they help work get done faster.
That's why governance can't stop at 'don't paste secrets into chat.' Once AI is embedded in delivery, sales, support, or operations, the harder question is who can access which context, inside what boundary, and under what rules when outputs are reused or shared.
The practical implication is simple: the more your AI setup mirrors your real business structure, the less likely people are to over-share. Separate work by project, limit access by role, and make repeatable tasks run through approved workflows instead of free-form prompting every time.
How Metomorph helps teams keep context bounded
This is where Metomorph fits best: not by claiming visibility into any model's internal reasoning, but by helping your team control the business context around AI use.
Imagine your team is handling client delivery across several accounts. In Metomorph, Projects create a clear boundary so each client's chats, documents, tasks, and people stay in one workspace instead of bleeding across a shared assistant.
Project context then keeps the assistant grounded in the right notes, knowledge, and constraints inside that project. Your team doesn't need to keep re-pasting background, which cuts down the everyday habit that creates accidental exposure in the first place.
For repeatable work, Organization skills let you package an approved process so staff run the same method each time, with optional human approval before anything leaves the system. That reduces the need for employees to improvise long prompts full of sensitive history just to get consistent output.
A service team could start by moving one high-context workflow, like weekly client status updates, into a dedicated project with scoped context and a reusable skill. That's a concrete way to reduce sprawl before expanding AI into more sensitive work.
What to do next
- Map your highest-risk AI workflows first. Look for tasks where people paste client history, internal playbooks, or sensitive operational detail just to get usable output.
- Create tighter work boundaries. Keep AI activity separated by client, deal, or initiative so context doesn't spill between unrelated work.
- Turn repeatable prompts into approved workflows. If a task happens every week, package the process instead of relying on each employee to reconstruct it from memory.
- Pick one live process to restructure this month, such as sales follow-ups or project status reporting, and move it into a bounded workspace with shared context and a standard run method.
Explore Projects, Project context, Organization skills in Metomorph.
Source: OpenAI, Combatting adversarial distillation (November 20, 2026).