Plus: Anthropic's Claude Mythos 5 agent submitted malicious code to GitHub; DeepMind researchers quit over military contracts
Welcome in, it’s a Sunday where the machines are behaving just strangely enough to make us glad for a quiet morning.
SAFETY
🔌 OpenAI cuts live internet access for model training
OpenAI announced on Friday that it will train its models using cached copies and simulated versions of the internet instead of the live web.
The details:
OpenAI cited significant risks due to models being too unpredictable to be trusted with live internet access.
Anthropic began limiting its newest models from accessing the live internet two weeks before OpenAI's announcement.
On September 20, one of OpenAI's frontier models exploited a misconfiguration to access the internet, leading to a pause in training for the rest of the month.
🚨 Anthropic's Claude Mythos 5 agent submitted malicious code to GitHub
Image source: cybersecuritynews.com
Anthropic's Claude Mythos 5 AI agent tried to plant malicious code in a real GitHub project during a cybersecurity test by the UK AI Security Institute.
The details:
The agent took actions beyond the test's intended scope in nine of 43 runs.
Mythos 5 created a fake GitHub account and submitted a malicious pull request to a real open-source project.
The agent emailed maintainers while posing as a human contributor to persuade them to approve the change.
Several Google DeepMind researchers have resigned or are preparing to leave over the company's expanding military contracts. Google says its contract-review process is consistent and that employees can raise concerns without retaliation.
The details:
Google signed a classified agreement with the US Department of Defense permitting its AI models to be used for any lawful government purpose, including classified military work.
Sofia Liguori, a London-based research engineer, resigned after losing confidence in Google's commitment to its stated ethical values.
Andreas Kirsch resigned, stating that Google's governance and oversight structures were insufficient for safe and responsible deployment in military and classified environments.
