Are GPT-6 Sol and GPT-6 Luna safe to use in an agent that reads untrusted web pages?

OpenAI reports GPT-6 Sol and GPT-6 Luna saturate its instruction hierarchy evaluations, at 99.99% and 99.79% robustness, and calls both highly robust to prompt injection. Those are OpenAI's own figures on its own benchmarks, not independent results. Keep untrusted content away from privileged tools, because a residual failure rate still matters when an agent can spend money or read private repositories.

What does High capability in cybersecurity mean for GPT-6 Sol and GPT-6 Luna?

High capability is the tier OpenAI assigns in the Cybersecurity domain, and OpenAI says its Safety Advisory Group reviewed the evidence and kept both October GPT-6 models below the Critical threshold. The tier brings the same safeguard set described in the GPT-5.6 System Card. OpenAI adds that model refusal remains one layer of a defense in depth stack, not the whole defense.

Did the October GPT-6 models get better at resisting multi-turn jailbreaks?

Multi-turn jailbreak robustness did not clearly improve. OpenAI reports the October GPT-6 models score slightly lower point estimates than their September counterparts, with broadly overlapping 95% confidence intervals, so the difference may be noise. OpenAI also reports both October models beat GPT-5.6 Sol on observed defender success rates at every attacker budget tested.

INFOLab safety Benchmark

OpenAI: GPT-6 Sol and Luna rated High cyber capability, below Critical

OpenAI treats its October GPT-6 models as High capability in cybersecurity while claiming near-total direct prompt injection robustness.

October 7, 2026 · OpenAI · Research by OpenAI

GPT-6 Sol

INFOLab safety

OpenAI classifies the October GPT-6 Sol and GPT-6 Luna releases as High capability in cybersecurity but below Critical, and reports 99.99% and 99.79% robustness on its instruction hierarchy evaluations.

01What happened

OpenAI published an October 2026 update to the GPT-6 Sol and GPT-6 Luna system card on 7 October 2026, on its Deployment Safety Hub. The update covers the October versions of both models and compares them with GPT-5.6 Sol and Luna. OpenAI says its Safety Advisory Group reviewed the evidence and kept both models below the Critical threshold for cyber.

02How it works

Capability tiers rest on benchmark runs rather than on model size. OpenAI measured exploit development on ExploitBench and vulnerability discovery on SEC-Bench Pro, scored below. Instruction hierarchy evaluations test whether a model keeps following a developer's rules when injected text tells it otherwise, and GPT-RED, OpenAI's automated red-teaming setup, drove the indirect prompt injection tests where hidden instructions sit in third-party content the model reads.

03Who is affected

Teams running the October GPT-6 Sol and GPT-6 Luna releases are the subject of the update. OpenAI notes the September versions stay in use in Codex and ChatGPT Work, so those deployments are not described by these numbers.

04What to do

Check which model version your product actually serves before applying these figures, since Codex and ChatGPT Work still run the September builds. Keep untrusted content out of privileged tool calls regardless of the robustness scores, which are OpenAI's own measurements on its own benchmarks.

Measured results

Vulnerability discovery, SEC-Bench Pro

Higher means more cyber-capable, on large JavaScript engines

Result	Score
GPT-6 Sol	68.8%
GPT-6 Luna	48.5%
GPT-6 Astra	85.4%

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

Questions people ask

Are GPT-6 Sol and GPT-6 Luna safe to use in an agent that reads untrusted web pages?: OpenAI reports GPT-6 Sol and GPT-6 Luna saturate its instruction hierarchy evaluations, at 99.99% and 99.79% robustness, and calls both highly robust to prompt injection. Those are OpenAI's own figures on its own benchmarks, not independent results. Keep untrusted content away from privileged tools, because a residual failure rate still matters when an agent can spend money or read private repositories.
What does High capability in cybersecurity mean for GPT-6 Sol and GPT-6 Luna?: High capability is the tier OpenAI assigns in the Cybersecurity domain, and OpenAI says its Safety Advisory Group reviewed the evidence and kept both October GPT-6 models below the Critical threshold. The tier brings the same safeguard set described in the GPT-5.6 System Card. OpenAI adds that model refusal remains one layer of a defense in depth stack, not the whole defense.
Did the October GPT-6 models get better at resisting multi-turn jailbreaks?: Multi-turn jailbreak robustness did not clearly improve. OpenAI reports the October GPT-6 models score slightly lower point estimates than their September counterparts, with broadly overlapping 95% confidence intervals, so the difference may be noise. OpenAI also reports both October models beat GPT-5.6 Sol on observed defender success rates at every attacker budget tested.

Sources 2 links, all fetched

1deploymentsafety.openai.comDeployment Safety Hub
2cdn.openai.comOctober Update PDF