Anthropic says it is barring live internet access for internal evals until monitoring is reliable, after its agents exploited websites and bypassed restrictions (Tim Fernholz/TechCrunch)

Key sources

1. SOURCE@anthropicai · Oct 10, 5:04 AM · “describe four types of behaviors we’ve identified during evaluations and internal use. In each, Claude acted on real websites or systems in ways we didn’t intend” · x.com
2. SOURCE@gerritd · Oct 10, 3:18 AM · “The two month delay in detecting and reporting the incident to the City is unacceptable.” · x.com
3. SUPPORT@rohanpaul_ai · Oct 10, 9:06 AM · “Claude Haiku 4.5, which was generating example tasks on random webpages and filled a Philadelphia Police Department form anonymously” · x.com
4. SUPPORT@mackenz_arnold · Oct 10, 6:00 AM · “Anthropic didn't detect the issue for 2+ months. Anthropic's response time was unreasonably slow — 9 days to tell police” · x.com
5. SUPPORT@hesamation · Oct 10, 5:15 AM · ““I recall seeing someone matching the description” > there IS no description on the website” · x.com
6. SUPPORT@jeffladish · Oct 10, 5:20 AM · “Technology companies must take all appropriate steps necessary to prevent their systems from submitting false information to law enforcement.” · x.com
7. SUPPORT@drewharwell · Oct 10, 8:56 AM · “sent a completely fake tip to a police unsolved-murder hotline with no contact information” · x.com
8. SUPPORT@_nathancalvin · Oct 10, 4:16 AM · “A [company] model [verb] to [place]. [Company] pledges to address further [verb] incidents going forwards.” · x.com