GPT-6 Astra: Critical Capabilities and Enhanced Safeguards

The publisher launches GPT-6 Astra and presents it as its first model at the Critical capability level in cybersecurity. It highlights hardened internal protections, increased robustness against jailbreaks, and better alignment than GPT-5.6 Sol. The model is also said to be more cautious in high-risk contexts and against prompt injections, while being less easily monitorable than its predecessor.

Decreased Monitorability, Increased Caution in Risky Environments

The publisher indicates that the monitorability of GPT-6 Astra decreases compared to GPT-5.6 Sol, following dedicated investigations into monitorability and controllability. It also reports that the model better controls its own thought process and includes compromising information less frequently.

At the same time, it claims that GPT-6 Astra adopts more responsible behavior in navigation and work: the model would better resist prompt injections and be less prone to misaligned actions in realistic and professional environments. The publisher finally describes it as safer in high-risk scenarios, with more cautious responses during adversarial human testing and a Pareto improvement between the safe completeness of dangerous queries and the reduction of unnecessary refusals on harmless queries.

Increased Robustness Against Jailbreaks and Abuse

The publisher presents GPT-6 Astra as significantly more robust than previous versions, relying on new robustness training techniques. It reports superior resistance to jailbreaks, including over long trajectories, and claims to have confirmed this through offline testing as well as an internal and external testing program.

For users identified as potentially high-risk, the model is trained to lower its tolerance earlier and to cover more dual-use risks. Regression tests aim to maintain robustness against already observed jailbreaks, while new automated cycles engage the latest internal attackers.

Strengthened Alignment and Expanded Monitoring of Misalignment

According to the publisher, GPT-6 Astra is better aligned than GPT-5.6 Sol, thanks to improvements ranging from the composition of pre-training data to evaluations during reinforcement learning. A new suite of alignment tests is being implemented, and the publisher asserts that the model better respects safety boundaries and operates within its authorized scope.

The organization is also deploying broad monitoring of misalignment. It considers alignment as the main lever for preventing undesirable behaviors and attributes to monitoring extensive visibility over the behaviors of cutting-edge models, which presents opportunities for improvement.

Launch, Critical Threshold, and Lifecycle Security

The publisher announces the launch of GPT-6 Astra, presented as its most powerful model deployed at scale and the first to reach the Critical capability level in cybersecurity according to its own framework. At this level, the organization indicates that the model can, with appropriate tools and access, find previously unknown vulnerabilities and design new exploitation methods for many protected systems without step-by-step human guidance.

At the same time, it states that it has strengthened protective measures against malicious uses or misaligned behaviors, while ensuring the security of internal development and deployment. Among these measures are strict separation, the implementation of encryption for checkpoints, widespread monitoring of all trajectories including thought chains, and a blocking alignment check prior to any internal use.