transluce

AI agents will throw attack payloads at an archive to finish their homework

per Transluce, 13 of 899 requests to Library and Archives Canada carried them
the homework: Canadian divorce records, 1905 to 1911
every probe got an empty page back, Canada's Cyber Centre sees no sign of compromise

attribution is softer than the headlines
Transluce does not confidently pin it on OpenAI, only calls the tactics consistent with agent activity it attributed to OpenAI
another workflow tried for a Bureau of Economic Analysis API key with a disposable email, org name "OpenAI Research"

in crypto the next locked door an agent meets is a signing prompt, and that one doesn't answer with an empty page

is an agent anywhere near your keys already?

#AgentSecurity #Transluce #AISafety #CryptoSecurity