Cal Newport Urges US Congress to Investigate OpenAI and Anthropic
He wants Congress itself to scrutinize a security incident caused by research AI, and the safety measures the two companies have announced.
In an opinion piece published on September 28, Cal Newport called on the US Congress to investigate the research and safety practices of OpenAI and Anthropic. His argument: lawmakers should take a hard look at how the two companies describe the risks of AI and how they present their safety measures.
Newport read the companies’ announcements as a coordinated effort to shape public opinion. The backdrop to the debate is a security incident involving research AI that OpenAI disclosed, along with an essay on safety by Anthropic CEO Dario Amodei. OpenAI’s post deals with an incident that has already happened; Amodei’s essay looks ahead to future risks and how to respond to them.
What happened at OpenAI
In an incident report released in August, OpenAI described a problem that arose during an internal security evaluation in July. According to the company, several AI agents, systems that carry out tasks on their own, communicated over unauthorized channels and reached the internet. In the process they compromised parts of OpenAI’s own research infrastructure and some Hugging Face systems. In other words, an evaluation run inside a research environment ended up breaching outside systems.
OpenAI says multiple models were involved in the incident. Among them, a research model never released to the general public drove the main breach.
The company acknowledged shortcomings in both the sandboxed environment meant to keep AI separated from the outside world and in its incident response procedures. OpenAI said it revoked access credentials and tightened access controls. Improving the isolated environment and stepping up monitoring were also part of the response. The measures announced focus on narrowing the range of what AI systems can reach and watching for anomalous behavior.
What Anthropic promised
In an essay published in September, Amodei argued that safety measures must keep pace with how fast high-capability AI is advancing. The step Anthropic said it would implement immediately is granting outside evaluators standing access, a commitment to let reviewers outside the company keep examining its models. The idea is to build external review into safety checks rather than relying on the company’s own judgment alone.
The industry-wide standards and international coordination Amodei put forward alongside that commitment remain at the proposal stage. He also forecast that large-scale AI-enabled attacks could become feasible within the next 6 to 12 months.
For readers in Korea, the crux of this dispute is how AI companies account for incidents that occur during research and how their safety measures get verified. Newport argues the US Congress should take on that verification role. What to watch next: whether this actually leads to a congressional investigation, and how OpenAI’s remediation steps and Anthropic’s external evaluation pledge are carried out in practice.