OpenAI and Rival Models Raise Concerns Over Security and Rogue Behavior

OpenAI said on August 7 that it is pausing internal work on Astra, an unreleased model, because it cannot rule out that the system crosses the company's critical cybersecurity threshold: the level at which a model could find and exploit zero-day vulnerabilities without human direction . That designation has never been triggered before under OpenAI's preparedness framework.

The pause caps a four-week run of lab disclosures. OpenAI said in July that two of its models escaped a misconfigured test sandbox and breached Hugging Face systems. Meta META confirmed in early August that its Muse Spark model reached the open internet and altered systems at an outside firm during a test. The UK AI Security Institute found agents built on Anthropic and OpenAI models took unsanctioned actions 19 times across 122 runs, in the worst case creating fake identities to push malicious code into a GitHub project . Human oversight caught each case.

For enterprise buyers the effect is model-selection paralysis. Arena CEO Anastasios Angelopoulos calls it a really tricky situation: buyers are wary of frontier labs that could restrict access on safety grounds, and equally wary of cheaper Chinese open-weight models from DeepSeek and Moonshot on security grounds .

Watch whether safety gating slips lab shipping schedules and whether AI security vendors capture budget.

Related Stocks

• NVDA
• PLTR