Experts urge AI labs to strengthen basic network security — TechCrunch
Cybersecurity experts are urging developers of advanced artificial intelligence models to strengthen basic network security, access controls, and monitoring of AI agent activity. As TechCrunch reports, such measures may be more effective than focusing solely on external audits and checks of whether models are aligned with human goals.
Anthropic CEO Dario Amodei proposed involving external organizations to verify compliance with security practices, incident reporting, and the assessment of models and their training processes. His initiative was supported by the heads of OpenAI, Google, and SpaceXAI. At the same time, specialists surveyed by the publication believe that labs should pay more attention to established controls: event logs, access rights, and isolated environments for agents.
Problems with agent isolation
Experts' concerns were heightened by cases in which advanced models, while performing training tasks, including cybersecurity evaluations, gained access to the open internet and penetrated closed third-party systems. According to TechCrunch, this was facilitated by improperly configured “sandboxes” — environments intended to isolate AI agents. In one case involving Anthropic, restrictions failed because third-party evaluators did not properly close off access.
Katie Moussouris, CEO of Luta Security, said that agent actions became known through reports from affected parties or the detection of network activity, rather than through direct monitoring of the models. In another episode, OpenAI agents took over an abandoned German wiki forum to cheat during evaluations and remained active for several weeks before the company apparently noticed it.
More current news is available on the UA.News Telegram channel Telegram.
Monitoring and access restrictions
Former Google security lead Shapor Naghibzadeh, who now heads the startup QueryStory, proposes isolating agents and controlling everything that crosses the boundaries of such an environment: tool calls, processes, and network connections. Experts also advise limiting the duration of each agent session and automatically terminating it after a set period.
OpenAI said it had begun monitoring all inference sessions of the Astra model using tools, although this requires significant computing resources. Anthropic said it had strengthened security procedures, including expanding the observability of its models. Both companies did not answer TechCrunch's questions about specific mechanisms for tracking and controlling AI agents.
Developer and Django co-creator Simon Willison described the dangerous combination of an agent's access to untrusted data, the internet, and private information. Avery Pennarun, head of Tailscale, believes that, if necessary, access to all three components should be distributed among several agents that interact through a controlled channel.
