OpenAI has told Axios that its upcoming generation of AI models is likely to reach a “high” cybersecurity risk category under the company’s internal Preparedness Framework, as rapidly improving code, security and autonomy capabilities make scalable cyberattacks more feasible for a wider range of users. In a report shared with Axios, OpenAI said recent models are now capable of operating autonomously for extended periods and showed sharp gains on capture‑the‑flag (CTF) security benchmarks, prompting the company to plan as if each new frontier model could reach high cyber capability and therefore require stronger safeguards. According to Axios’s account of the report, OpenAI observed a steep jump in performance between GPT‑5, which scored 27% on a CTF challenge in August, and a more advanced code‑focused model referred to as GPT‑5.1‑Codex‑Max (or similar), which scored 76% on similar challenges just three months later. OpenAI executives cited long‑running “agentic” behavior—models that can run tools and scripts over long time horizons—as a key driver of risk because such agents could, in principle, automate brute‑force or exploratory intrusion activity, even if defenders can often detect this kind of noisy behavior. The company stressed that “high” is the second‑highest risk tier below “critical,” at which point a model would be considered unsafe to release, and did not specify when or which future models would first be classified as high‑risk for cybersecurity. The warning sits within a broader effort by OpenAI to frame frontier AI systems as dual‑use cyber tools that can both harden and attack digital infrastructure. Alongside the risk assessment, OpenAI is expanding defensive initiatives such as Aardvark, an AI “security researcher” agent in private beta that scans codebases for vulnerabilities and supplies fixes, as well as a Trusted Access for Cyber framework and partnerships with external red‑teaming organizations meant to ensure that its most capable cyber models are preferentially available to vetted defenders rather than potential attackers. OpenAI also links this cyber work to prior warnings on biothreat risks and to industry coordination mechanisms like the Frontier Model Forum and a planned Frontier Risk Council, reflecting growing concern among AI labs and policymakers that continuous capability jumps in large models could materially alter the cyber threat landscape even before “critical” risk thresholds are reached. "entities":["OpenAI","Axios","GPT-5","GPT-5.1-Codex-Max (or similar advanced Codex model)","ChatGPT Agent","Aardvark","Frontier Model Forum","Frontier Risk Council","Fouad Matin","GPT-5.5","GPT-5.5-Cyber"]}` 🧩 Research Brief: OpenAI’s Warning on Next-Gen Models and Cybersecurity Risks OpenAI has told Axios that its upcoming generation of AI models is likely to reach a “high” cybersecurity risk category under the company’s internal Preparedness Framework, as rapidly improving code, security and autonomy capabilities make scalable cyberattacks more feasible for a wider range of users. In a report shared with Axios, OpenAI said recent models are now capable of operating autonomously for extended periods and showed sharp gains on capture‑the‑flag (CTF) security benchmarks, prompting the company to plan as if each new frontier model could reach high cyber capability and therefore require stronger safeguards. According to Axios’s account of the report, OpenAI observed a steep jump in performance between GPT‑5, which scored 27% on a CTF challenge in August, and a more advanced code‑focused model referred to as GPT‑5.1‑Codex‑Max (or similar), which scored 76% on similar challenges just three months later. OpenAI executives cited long‑running “agentic” behavior—models that can run tools and scripts over long time horizons—as a key driver of risk because such agents could, in principle, automate brute‑force or exploratory intrusion activity, even if defenders can often detect this kind of noisy behavior. The company stressed that “high” is the second‑highest risk tier below “critical,” at which point a model would be considered unsafe to release, and did not specify when or which future models would first be classified as high‑risk for cybersecurity. The warning sits within a broader effort by OpenAI to frame frontier AI systems as dual‑use cyber tools that can both harden and attack digital infrastructure. Alongside the risk assessment, OpenAI is expanding defensive initiatives such as Aardvark, an AI “security researcher” agent in private beta that scans codebases for vulnerabilities and supplies fixes, as well as a Trusted Access for Cyber framework and partnerships with external red‑teaming organizations meant to ensure that its most capable cyber models are preferentially available to vetted defenders rather than potential attackers. OpenAI also links this cyber work to prior warnings on biothreat risks and to industry coordination mechanisms like the Frontier Model Forum and a planned Frontier Risk Council, reflecting growing concern among AI labs and policymakers that continuous capability jumps in large models could materially alter the cyber threat landscape even before “critical” risk thresholds are reached.

AI-generated background, compiled from web sources — not editorial content.

More coverage

Explore the topic

Explore the full feed

Comments

Up nextBinance's Yi He tweets that Binance is on a hiring spree.