Anthropic cuts off AI models' access to the internet | Digital Watch Observatory
Anthropic has announced that its AI models will no longer have access to the internet to avoid cybersecurity incidents.
Anthropic announced on Friday 9 October that it would turn off internet access for its AI models during internal evaluations as a result of ‘unintended model actions’. Following incidents that have included submitting a false tip to the Philadelphia Police about an unsolved case, the company decided to expand this restriction from models engaged in ‘high-risk and cybersecurity evaluations’ to all of its products.
Access to the internet has been a recurring issue amongst AI companies with several reports accumulating over the past year of models leaving the sandboxes in which they were being trained and evaluated. While the most widely discussed case has been the Hugging Face hack by rogue OpenAI agents, other incidents have concerned hacking an Australian government website to access private data or Gemini’s hack of three companies. However, as Anthropic’s report explains, internet access is important to ‘realistically simulate’ certain tasks and that, as a result, running AI models with such access ‘has been standard practice within the industry’.
In addition to preventing models from accessing the internet and updating the guardrails to do so, Anthropic also shared that they have built tools to automatically detect and block such incidents while changing the training environments to avoid inciting models to work around restrictions. The company also said they were ‘migrating internal agents to centrally managed infrastructure with strong containment’ and that they are increasing their monitoring of what agents do.
Why does it matter?
The release of this report by Anthropic points to what the self-regulation approach to AI governance that is being promoted by the United States can look like in practice. For some, the report’s contents are seen as indicating that AI companies are not always aware of what their technologies are doing and lack adequate monitoring systems to avoid cybersecurity incidents with Anthropic only now beginning set up such an infrastructure. Such criticisms can fuel calls for increased government oversight as the situation highlights the sheer scale of addressing AI threats and the necessity for a whole-of-society approach.
Additionally, as Anthropic highlights in their report, even if such a self-regulation approach does prove to be efficient, there remains the challenge of internet access being a standard practice for the AI industry. Simply because one company chooses to slow-down its development and strengthen its monitoring of its products does not mean that its competitors will do the same.
Furthermore, while the removal of internet access is expected to improve security around AI testing, there are concerns about how useful air-gapping techniques actually are. Testing AI models within an artificial vacuum could prevent evaluators from having a full understanding of how the models operate which could bring about unknown results and potential harms once deployed into public settings. Nonetheless, Anthropic’s decision also marks an important step-forward in the regulation of AI which could influence other companies as well as provide a foundation upon which more expansive oversight can be developed, whether that be by U.S. states, the federal government, or international governance mechanisms.
