Understand this story
On October 31, 2026, Nathan Lambert offered a judgment, arguing that the AI community focuses on "which types of models cause harm" while neglecting how cybersecurity maintains balance, who is carrying out the harm, and how they use new tools. In an article the same day, he reflected on his earlier judgments about the cyber risks of open-weight models: more than a month after the release of GLM-5.3, public evidence did not show a step-change in cyber risk; if Claude Mythos were accidentally leaked as open weights, the world would most likely just see risk accelerate rather than catastrophe. He mentioned that Epoch AI used GLM-5.3 to find a vulnerability in Firefox that could silently steal files, showing that exploiting a vulnerability still requires coordination across multiple steps, and acknowledged that he should pay more attention to the reasonable warnings from the AI safety community about a new phase of cyber instability. He also believes that open-source models will be used for cyberattacks just like closed-source models, but they are not a super-leverage tool that could cause cybercrime to surge 100-fold; the real bottleneck lies elsewhere.
AI-generated from public reports
Latest development · Report
Nathan Lambert: Open Models Won't Cause a Hundredfold Surge in CybercrimeDevelopments
1 confirmed developments
Reports and discussion timeline
Browse reports and discussion by time. Their evidence levels remain distinct.
Oct 31
Nathan Lambert
Nathan Lambert: Open Models Won't Cause a Hundredfold Surge in Cybercrime
Nathan Lambert believes that open models will be used for cyberattacks just like closed models, but they are not a super-leverage tool capable of causing a 100-fold surge in cybercrime—the real bottleneck lies elsewhere. He also criticizes the AI community for focusing on "which type of model causes harm" while ignoring how cybersecurity maintains an equilibrium, who is carrying out the harm, and how they use new tools.
Nathan Lambert
Nathan Lambert responds to questions about cybersecurity risks of open-source models
Nathan Lambert said he does not believe his previous views on open-source models and cybersecurity risks were wrong. He believes the AI community has been looking at this from the wrong perspective: rather than asking which types of models are causing harm, it is more important to focus on how cybersecurity maintains balance, who is carrying out the harm, and how they use new tools.
Interconnects
Was I wrong about open models and cybersecurity risk?
Nathan Lambert reflects on his earlier judgment about the cybersecurity risks of open-weight models: more than a month after the release of GLM-5.3, public evidence does not show a step-change in cyber risk; if Claude Mythos were accidentally leaked as open weights, the world would most likely see an acceleration of risk rather than catastrophe. Epoch AI used GLM-5.3 to find a vulnerability in Firefox that could silently exfiltrate files, showing that exploiting a vulnerability still requires multiple coordinated steps. He acknowledges that he should have taken more seriously the reasonable warnings from the AI safety community about a new phase of cyber instability.
Heat trend for this story
- Current
- 5
- Period peak
- 19 Oct 31 03:00
- 24-hour change
- -50%
Move the pointer or select the chart to inspect hourly heat. Use the left and right arrow keys with a keyboard.