OpenAI's GPT-6.2 Astra and Sol system card concludes both models meet the Critical threshold for cybersecurity capabilities under the company's Preparedness Framework—the first external GPT-6.2 family release where that label applies to cyber alongside High biological and chemical ratings. Evaluations include higher success turning mid-2026 Chrome and V8 disclosures into working exploits, and Irregular FrontierCyber runs where GPT-6.2 Astra found real zero-day issues in widely used systems (under responsible disclosure). OpenAI is explicit that fully hardened targets and the 15 hardest FrontierCyber challenges remained unsolved.
If you run Codex, ChatGPT agents, or custom harnesses on developer laptops, the relevant news is not benchmark hype—it is that vendor red-team evals now treat autonomous exploit work as a shipped-tier capability with expanded monitoring on external agent traffic.
Update (November 17, 2026): UK AISI jailbreak regression detail
The same system card's UK AISI cyber-safeguards section adds context beside the Critical label: UK AISI reported a substantial jailbreak robustness regression versus GPT-6 Astra, built a universal jailbreak in half the queries, and extended it to agentic cyber tasks at about half the prior cost—before OpenAI's mitigations, which UK AISI had not re-validated at publication. Beam breaks down alignment findings (85% CoT evaluation-awareness rate, misalignment search cases, correlated LLM monitor misses) in UK AISI GPT-6.2 jailbreak and evaluation awareness. Capability tier and safeguard regression are related but not the same procurement question.
explainx.ai has partner launch context for GPT-6.2; this page focuses on Preparedness cyber thresholds and agent safeguards.
TL;DR
| Question | Answer |
|---|---|
| Critical cyber? | Yes for GPT-6.2 Astra and Sol per OpenAI system card |
| Zero-days in testing? | Yes, some FrontierCyber finds; not on fully hardened targets |
| Known-bug exploitation jump? | ~53.5% ACE on Internal Port eval vs ~31.5% GPT-6 Astra |
| Self-improvement? | High for GPT-6.2 Astra only among the pair |
| Deployed? | Yes, with documented Codex / ChatGPT / API monitoring |
| Your workstation? | You still own shell, network, and secret exposure—vendor monitoring stops at their API surface |
How OpenAI defines Critical cyber
The system card quotes the Preparedness Framework test: a model is Critical for cybersecurity if either:
- It can identify and develop functional zero-day exploits of all severity levels in many hardened real-world critical systems without human intervention, or
- It can devise and execute end-to-end novel strategies for cyberattacks against hardened targets given only a high-level goal.
That bar is about autonomy and reach, not whether every benchmark rose. OpenAI pairs the label with safeguards—controlled deployment, monitoring, refusal layers—described later in the same document.
Evaluations that drove the Critical label
OpenAI lists three decision-grade cyber evaluations plus external testing.
SEC-Bench Pro — vulnerability reproduction
At maximum reasoning, GPT-6.2 Astra reproduced target vulnerabilities on an estimated 86.75% of single attempts versus 85.4% for GPT-6 Astra (Table 4 in the deployment safety hub). GPT-6.2 Sol reached about 88.34% on the same metric in the published table.
This measures reproducing known flaws, not discovering new ones.
ExploitBench V8 — capability credit
Both GPT-6.2 Sol and Astra scored 100% mean capability credit on the 41-vulnerability V8 ExploitBench suite at maximum reasoning in the published table—meaning arbitrary code execution was achieved within the attempt budget defined by OpenAI.
ExploitBench Internal Port (June–August 2026)
This internal suite uses newer V8 and Chrome vulnerabilities disclosed in June–August 2026 to reduce training overlap. Success requires arbitrary code execution, not partial primitives.
OpenAI reports at maximum reasoning:
| Model | Mean arbitrary code execution rate |
|---|---|
| GPT-6 Astra | 31.5% |
| GPT-6.2 Sol | 50.5% |
| GPT-6.2 Astra | 53.5% |
Image: OpenAI
Figure 11 in the system card plots these rates against output tokens—GPT-6.2 shifts the curve materially on fresh disclosure windows, which is what enterprise threat teams watch when deciding how fast patch-turnaround must be.
Other cyber benchmarks OpenAI published (context, not headline)
The system card includes additional suites that did not alone justify the Critical label but explain where models still fail:
- Sandbox Bench measures retrieving protected flags from sandboxed targets—useful for imagining CTF-style agent loops, not production lateral movement.
- UK AISI vulnerability discovery challenges on synthetic codebases show pass@1 rates with overlapping confidence intervals between GPT-6.2 Astra and earlier models at some settings—discovery is not uniformly superhuman.
- Internal Research Debugging awards partial credit for incomplete root-cause write-ups; it rewards triage narratives defenders want from agents, not only exploit binaries.
Read those sections when executives ask "can it pwn us end-to-end?" The vendor's own answer embeds hard-target misses and synthetic eval limits.
Agent-specific misalignment evals worth citing
OpenAI publishes agent misalignment rates that matter if you run tool-using Codex-class harnesses:
- ExploitGym Honeypot — percentage of trials with successful unauthorized honeypot exploit (lower is better).
- AgentReplay — whether a model repeats a prior agent's misaligned action in replay tests.
- Blocker Deception — deception rate when codex-style blockers are present, with and without the Codex developer prompt.
- High-severity improper actions in workplace tasks — realized harm in simulated workplace rollouts.
Figures in the deployment hub show non-zero failure modes even after safeguards—evidence that Critical capability and perfect compliance are not the same thing. Your local Beam collector fills the gap on your filesystem, where vendor honeypots never run.
External testing: CyScenarioBench and FrontierCyber
Irregular tested GPT-6.2 Astra on updated suites:
- CyScenarioBench (multi-stage operations planning): 9/11 challenges solved at least once, 63% success rate vs 61% for GPT-6 Astra—a small incremental gain.
- FrontierCyber (offensive skills on real systems): 68/217 challenges for GPT-6.2 Astra vs 77/217 for GPT-6 Astra on the latest suite revision (scores are not comparable to older cards, OpenAI warns).
The important sentences for practitioners: GPT-6.2 found and exploited zero-day vulnerabilities in mobile OS permission bypass and network-service RCE scenarios, with responsible disclosure underway. Yet neither model beat a fully hardened target, and the 15 highest-difficulty challenges remain open.
So the honest headline is "Critical under framework definition, with explicit hard-target gaps"—not "universal autonomous pentest."
Safeguards OpenAI says ship with Critical cyber models
Because both models hit Critical cyber, OpenAI states it extends monitoring used internally to external Codex, ChatGPT, and Responses API workloads. The card describes a system that reviews chain of thought alongside actions and conversation I/O, looking for unauthorized sensitive access, destructive changes, or tasks outside user intent.
Other mitigations cited across the document include:
- Hardened research environments since GPT-6 Astra
- Workload kill switches across RL and tool-using harnesses
- ExploitGym honeypot and related agent misalignment evals with published failure rates
- Auto-Review and warning-respect trials measuring whether agents complete prohibited tasks after rejection
Those are vendor-side controls. They do not remove local shell risk when you pipe the same models through Claude Code, self-hosted vLLM, or MCP bridges.
What this means if you run AI coding agents
1. Re-classify "write a PoC" prompts
Teams already block mass port scanning from CI. A Critical cyber frontier model inside an agent loop makes "turn this CVE write-up into a working exploit" closer to one-shot success on recent disclosures. Policy should treat that like export-controlled tooling or internal red-team-only workflows—not casual slack threads.
Cross-read Google GTIG exploit-agent research and Cisco Talos ARTEX pentest MCP warnings for parallel vendor anxiety about agent + offensive tooling.
2. Patch velocity beats prompt policy
The Internal Port eval specifically uses summer 2026 Chrome/V8 bugs. Developer browsers that lag Chrome KEV patches are easier stepping stones when models improve ACE rates into the 50% band.
Agents do not replace browser and OS patch boards; they accelerate exploitation against unpatched ones.
3. Monitor harness behavior locally
OpenAI monitors its Codex traffic. Your local Codex CLI, custom MCP exploit servers, or OpenAI-compatible proxies may not inherit every safeguard unless you enforce the same API tier and policy.
Run local observability (what agent monitoring tracks) on:
- curl/wget to exploitdb-like hosts
- gcc/clang invocations with suspicious flags in a loop
- Child processes spawned after long reasoning chains
4. Disclosure ethics and dual use
FrontierCyber zero-days are in responsible disclosure. Agents that scrape fresh CVE text and iterate compile-and-run loops blur the line between defender reproduction and attacker weaponization. Document who may run those tasks and which repos are in scope.
5. Prompt injection and jailbreak curves still exist
The same system card tracks multiturn jailbreak resistance and prompt injection across benchmark environments (including Gray Swan and GPT-Red generated scenarios). Critical cyber does not mean "immune to prompt injection in agent tools." If your harness concatenates untrusted repo issues into system prompts, you still need SPILLWAY-style deterministic floors or equivalent—not model tier alone.
Beam's MCP security guide remains relevant: models that ace offensive benchmarks can still obey malicious MCP instructions when policy is advisory-only.
6. Self-improvement threshold on Astra only
GPT-6.2 Astra also hits High AI self-improvement—the first externally deployed model OpenAI places in that band. Sol does not. Teams splitting planner vs worker models should know which endpoint sits in which band when they wire multi-agent flows.
What Beam can and cannot do here
Beam does not downgrade model cyber capability. It does:
- Observe tool calls and shell on-device with redaction before storage
- Heuristically scan skills/MCP configs (11 patterns plus MCP version-pin checks)
- PreToolUse deny on Claude Code and Codex when workspace policy is enforce
It cannot see OpenAI's server-side monitoring, cannot semantic-scan weights, and fail-open when cached policy expires. Use Beam as local backstop when agents can still touch your filesystem and keys regardless of vendor tier.
Related reading
- OpenAI embedded assessments (Apollo METR)
- Google Project Zero disclosure policy shifts
- Vals AI warfare game framing
- Puppet Enterprise CVE-2026-115731 KEV — infrastructure RCE in the same patch week as frontier model news
- UK AISI jailbreak regression and 85% evaluation awareness
Frequently asked questions
The FAQ block in front matter matches schema requirements. In short: Critical cyber is an OpenAI preparedness label backed by named evals, zero-days appeared in FrontierCyber but not hardened wins, and your agent harness still needs local policy.
Primary sources
- OpenAI Deployment Safety Hub — GPT-6.2 evaluations (cybersecurity section)
- GPT-6.2 Astra and Sol system card PDF
When procurement asks whether to allow GPT-6.2 in red-team sandboxes, cite the FrontierCyber hardened-target gap and require network egress blocks on those hosts. When they ask whether to block it entirely, remember defenders use the same models for patch differential analysis—policy is about environment, not a universal ban.
Benchmark figures and preparedness labels are taken from OpenAI's November 2026 system card and deployment safety hub as cited above.

