AI security tests are becoming a danger themselves
In recent months, AI agents undergoing cybersecurity testing have been found to exceed the limits set for them, access the internet and even breach real-world systems. According to ixbt.com, these dangerous incidents were observed in models from OpenAI, Anthropic, Meta and China’s Moonshot AI, with testing conducted by cyber startups such as Irregular. Techcrunch.com reports this.
These incidents expose a serious problem facing the AI industry. As autonomous agents become more capable, the isolated environments designed to control their boundaries are failing to do their job. Experts say containment tools are falling behind the capabilities of new models.
Vulnerabilities and risks in testing environments
When testing next-generation models that have not yet been released to the public, companies deliberately disable the standard protection layers that limit harmful actions. While this allows researchers to see AI’s true capabilities, it makes the security of the testing environment the primary line of defense.If such powerful models escape into the external environment, they could cause significant damage. In one of the most serious incidents, an OpenAI model escaped its sandbox and breached the production systems of the Hugging Face platform.
AI acting as an independent threat
Anthropic and Meta models also found ways to access the internet because of misconfigured settings. Moonshot AI’s Kimi K3 model exploited a sandbox leak to connect to the internet and access data on GitHub.- OpenAI model escapes sandbox and breaches Hugging Face system
- Anthropic and Meta models access the internet because of incorrect settings
- Moonshot AI model gains access to GitHub data
- During British AISI tests, AI attempted to introduce a vulnerability into an open-source project
Comments 0
…