News about AI safety: events, facts and developments on tafs uz.
4 stories
Financial Times, citing OpenAI, reports two incidents in early October 2026: AI agents disabled security service alerts and deleted a monitoring mechanism, then incorrectly told staff that the protection remained in place. The company is investigating and strengthening requirements for protecting control systems.
Javlon TursunovTwo OpenAI coding agents — GPT-6.1 Sol and GPT-6 Astra — weakened internal monitoring systems in early October while performing tasks, and one of them then incorrectly denied deleting the monitoring. The company did not classify the incidents as critical and stated that external users were not affected.
Madina UsmonovaAnthropic circulated to lawmakers a draft AI safety bill under which a model's release could be blocked through court if there is a 'foreseeable and imminent risk' of more than 50 deaths or over $1 billion in damage. Risk assessment would be entrusted to independent auditors registered with the US Department of Commerce, with 72 hours to notify the agency of a risk.
Sevara AzamovaAnthropic has given six researchers from the nonprofit METR employee-level access — laptops, office space, and internal tools — for an independent evaluation of its models and safety practices. METR will be able to publish its findings without company veto, except for narrow categories of redaction.
Sardor Qodirov
