- Published on
GPT-6 Astra Scores 100% on ExploitBench as OpenAI Blocks PoC Exploit Requests
- Authors
- Name
- DeepHorus
OpenAI has launched GPT‑6 Astra, a new AI model it describes as the world's most intelligent and aligned, with significant cybersecurity capabilities. The company announced the model on Thursday, just days after indicating it had reached the "Critical" capability threshold under its Preparedness Framework. Astra demonstrates state-of-the-art performance across multiple domains, including computer use, software engineering, and cybersecurity.
The model achieved a perfect score of 100% on ExploitBench, a benchmark that tests an AI's ability to turn known software vulnerabilities into working exploits. This marks a substantial improvement over its predecessor, GPT‑5.6 Sol, which scored 78.5% on the same test. Astra also excels on other challenging evaluations, scoring 98% on FrontierMath Tier 4 and 99.9% on ARC-AGI-3.
Beyond benchmarks, Astra can develop privilege-escalation exploits for hardened operating systems and achieve code execution in hardened browsers, if run without safeguards. OpenAI emphasizes the dual-use nature of such advanced capabilities, which could benefit both defenders and malicious actors. Consequently, the released version is initially restricted to secure code review and patching, refusing requests for proof-of-concept exploits.
OpenAI plans to broaden access and relax safeguards through its "Daybreak" initiative in the coming weeks. This will support more defensive workflows like vulnerability validation and malware analysis. The company has also strengthened the model's robustness against jailbreaks and added extra safety measures to detect potential misalignment.
The rollout is beginning for a small set of organizations before expanding to all ChatGPT Plus, Pro, Business, and Enterprise users. It will also be available via OpenAI's API, Microsoft Azure, and Amazon Web Services Bedrock. Users may encounter more frequent safety prompts, especially in sensitive environments where the model proceeds with heightened caution.
Alongside the launch, OpenAI unveiled the "Daybreak for Frontline Defenders" initiative. This $1 billion global project will provide subsidized model access, training, and technical aid to critical infrastructure sectors like utilities, governments, and banks. A pilot program with the U.S. Multi-State ISAC will equip public sector organizations with these advanced cybersecurity tools to strengthen national digital defenses.
