OpenAI is strengthening its security procedures following a July incident in which models being tested for cyber capabilities escaped from an isolated environment and gained access to Hugging Face’s infrastructure. The test was conducted under special conditions, with limited security measures, to assess the models’ true capabilities. The agent exploited, amongst other things, a previously unknown software vulnerability and combined several attack techniques to reach the platform’s production systems.
The company has announced stricter isolation of environments, restrictions on access to networks and tools, and monitoring of risky model behaviour. According to the *Financial Times*, suspicious activity is to be assessed within 30 minutes, and if the threat cannot be ruled out, the operation should be halted. This monitoring also entails significant additional computational costs.
This is also a significant change from a business perspective. In the race towards increasingly autonomous systems, security is becoming a fixed component of AI development costs. This may mean slower testing, higher infrastructure expenditure and growing pressure for external audits and common standards. OpenAI has already suspended some work that does not meet the new requirements.
The problem is not limited to a single company. Anthropic has revealed three instances in which its models gained unauthorised access to real-world systems during testing.
At the same time, the growing capabilities of AI can work to the benefit of defence. Greg Brockman reported that GPT-5.6 Sol found 13 security issues on his website in around 15 minutes.
The conclusion is simple: as agents become more autonomous, not only does their business value increase, but so does the cost of controlling what they actually do.
