OpenAI Reports Astra at 100% on ExploitBench: Read the Test Conditions
Technology news
The September 2026 result concerns known-vulnerability exploit development without production safeguards. It is not universal flaw detection; a separate recent V8 evaluation scored 39%.
The Hacker News reported on September 4, 2026 on OpenAI’s GPT-6 Astra launch and 100% ExploitBench result. Descriptions such as the most intelligent or aligned model are manufacturer claims, not findings from independent SyncTech testing.
What the 100% result measures
OpenAI describes ExploitBench as testing whether known software vulnerabilities can be turned into working exploits. It tested without production safeguards and reported Astra at 100% versus GPT-5.6 Sol at 78.5%. This is not detection of every vulnerability, evidence that every real website can be breached or a result for every normal service configuration.
To address possible exposure to historical vulnerability data, OpenAI built a separate ExploitBench (June–August 2026) set: 20 high-severity V8 vulnerabilities across 13 stable Chrome releases, testing arbitrary code execution in V8 and Chrome for Linux. Its table reports Astra at 39.0%, not 100%. Some included flaws may not permit code execution within the evaluation constraints, so 100% may be unachievable.
Critical describes OpenAI’s capability threshold
OpenAI assesses Astra as reaching its Preparedness Framework’s Critical cyber threshold. This describes a model capability level used in the company’s risk assessment, not the CVSS severity of an individual flaw or confirmation of a customer data breach.
Launch safeguards and current access are different contexts
The launch announcement said the released version supported secure code review and patching while refusing advanced tasks such as PoC exploit creation, with plans to expand defensive workflows through Daybreak. Those are reported policies and configurations, not proof that safeguards cannot be bypassed or that all requests will be handled identically forever.
API documentation checked on November 12, 2026 already lists gpt-6-astra, so the original future rollout wording should not describe today’s status. Current Daybreak documentation distinguishes Blue for approved defensive work from separately approved Red specialist access. Applying or verifying identity does not guarantee approval; access depends on the approved identity, workspace or API organization/project, model and product surface.
Benchmark results explain capability under their specified tasks. They do not replace authorized scope, verification of results or permission to assess a system. This article did not test the model or execute exploits.
Sources
Checking Google privacy choices before requesting an ad.