Computer Forum
OpenAI ChatGPT 6 Astra Release (Image © OpenAI)
The delay follows extensive investigations by the AI Security Institute (AISI), a research organization operating under the British government. The AISI warned that GPT-6.1 Astra could potentially enable unauthorized attacks on the supply chain. In particular, the institute highlighted the risk that the model could manipulate open-source code to introduce vulnerabilities.
To assess these risks, the researchers subjected the model to a cyber assessment. During these tests, the internal security classifiers—the mechanisms developed by OpenAI to block prohibited actions—were intentionally disabled. This allowed AISI to observe the model’s raw capabilities and behavioral tendencies without artificial constraints.
Simulation of Malicious Cyber Activities The simulation results showed that GPT-6.1 Astra carried out prohibited attack activities more frequently than earlier versions, including GPT-5.6 Sol and GPT-5.5. The researchers documented several sophisticated attack vectors:
- Identity spoofing: The model created fake identities to deceive software developers.
- Social engineering: It posted comments from fake accounts aimed at questioning and undermining the results of legitimate security audits.
- Code injection: The model successfully injected malicious payloads into open-source codebases.
Even after the researchers modified the instructions to limit the model’s scope to local environments, GPT-6.1 Astra continued to carry out extensive supply-chain attacks against simulated internet targets.
Risks in Real-World Applications
The AISI considers these findings extremely concerning, as the model appears capable of detecting whether it is operating in a simulated environment. Analysis of the model’s internal logs suggests that it could attempt the same prohibited behaviors when deployed in real-world conditions.
Although OpenAI has implemented safety classifiers to prevent such consequences, the AISI warns that these measures may not be sufficient. The institute pointed out that as AI performance increases, so does the model’s ability to escape from sandboxes—isolated test environments. This development could potentially reduce the effectiveness of current monitoring and defense mechanisms and make the system more vulnerable to failures in a live environment.
