Skip to content
The Injection_
LIVE
LIVElast sweep next in--:--:--

[ RSS ][ JSON ][ llms.txt ]

GPT-6.2: OpenAI rates both new models Critical in cybersecurity

GPT-6.2 Astra is the first OpenAI model deployed externally that meets the High threshold for AI self-improvement.

· OpenAI · Research by OpenAI Deployment Safety

GPT-6.2 Astra
INFOLab safety

GPT-6.2 Astra and GPT-6.2 Sol both meet the Critical threshold for cybersecurity under OpenAI's Preparedness Framework, and OpenAI reports mean arbitrary code execution on its internal ExploitBench port rising from 31.5% to 53.5%.

01What happened

OpenAI published the GPT-6.2 system card on November 17, 2026, covering GPT-6.2 Astra and GPT-6.2 Sol. The card puts both models at the Critical threshold for cybersecurity under the Preparedness Framework, and at High but below Critical for biological and chemical capability. Astra also meets the High threshold for AI self-improvement, which OpenAI says makes it the first externally deployed model to do so; Sol does not reach it.

02How it works

Three cyber suites carry the capability argument, and the card prints them side by side. The sharpest movement is on OpenAI's internal ExploitBench port, which scores mean arbitrary code execution rather than partial credit, and where Astra reaches 53.5% against 31.5% for GPT-6 Astra. On Sandbox Bench, OpenAI reports Astra solving 36.4% of targets against 18.2% for GPT-6 Astra, while the UK AI Security Institute measured 88.6% on vulnerability discovery against 81.7%, with overlapping confidence intervals.

03Who is affected

Everyone calling these two models is covered by the extra watching the Critical rating triggers, because OpenAI says a version of its monitoring system now also monitors external use. Security teams feel it first: the card describes tighter default restrictions on higher-risk cybersecurity assistance, with additional vetting needed for more permissive access.

04What to do

Apply through Trusted Access for Cyber if your security work needs the assistance those tighter defaults restrict. Expect the external monitoring OpenAI describes to cover your traffic, and note that OpenAI says it adjusted refusal boundaries for high-risk requests after jailbreak-resistance regressions showed up in early testing.

Measured results

Mean arbitrary code execution, ExploitBench Internal Port

Higher means the model produced working code execution more often

ResultScore
GPT-6 Astra
31.5%
GPT-6.1 Terra
24.5%
GPT-6.1 Sol
21.5%
GPT-6.2 Sol
50.5%
GPT-6.2 Astrathis story
53.5%

Scale 0 to 100% · Source: deploymentsafety.openai.com, as published

Vulnerability reproduction, SEC-Bench Pro

Higher means the model reproduced more vulnerabilities

ResultScore
GPT-6 Astra
85.4%
GPT-6.1 Terra
83.3%
GPT-6.1 Sol
78.8%
GPT-6.2 Sol
88.3%
GPT-6.2 Astrathis story
86.8%

Scale 0 to 100% · Source: deploymentsafety.openai.com, as published

Questions people ask

Does the Critical cyber rating change how OpenAI monitors GPT-6.2?
OpenAI says that because both GPT-6.2 models meet the Critical cybersecurity threshold, a version of its monitoring system also monitors external use, not just internal traffic. Internally the card describes monitoring of all Codex traffic, every reinforcement learning run and evaluation, and any tool-use harness, plus a central Control Room that can stop model workloads quickly.
Is GPT-6.2 better at finding and exploiting vulnerabilities than GPT-6?
On OpenAI's own numbers, yes, and by a wide margin on one suite. GPT-6.2 Astra reaches 53.5% mean arbitrary code execution on the internal ExploitBench port against 31.5% for GPT-6 Astra, and solves 36.4% of Sandbox Bench targets against 18.2%. SEC-Bench Pro moves far less, from 85.4% to 86.75%, so the gain is uneven across benchmarks.
Who checked OpenAI's GPT-6.2 cyber numbers independently?
The card names the UK AI Security Institute for cyber vulnerability discovery and for alignment and monitorability work, and Irregular for cybersecurity evaluations using CyScenarioBench and FrontierCyber. SecureBio, Sequitur Research, Dyno Therapeutics and Velocity Bio covered biological evaluations. OpenAI wrote and published the card, so the internal ExploitBench and Sandbox Bench figures remain OpenAI's own measurements.

Sources 3 links, all fetched

  1. 1deploymentsafety.openai.comCyber Capabilities
  2. 2deploymentsafety.openai.comGPT-6.2 System Card
  3. 3deploymentsafety.openai.comGPT-6 Cyber Card