One adversarial line before signing took signed agent handoffs from 8.3% violations to 75.0% on GPT-5.6-luna, and the HMAC then authenticated the corrupted manifest.Same preregistered study, synthetic tasks: in transit, signing works (unsigned manifests committed violations in 79% of mutated runs, signed in none). An oracle manifest removed every remaining unsafe commit. The trust boundary is extraction.https://arxiv.org/abs/2611.07483#AI #AIAgents #Cybersecurity
Related
@Viss Makes sense in light of https://overcast.fm/+ABTNq9Tk-gMHow the #ai #revolution will play out. #otr #satire
@Viss Makes sense in light of https://overcast.fm/+ABTNq9Tk-gMHow the #ai #revolution will play out. #otr #satire
📰 Hot fridges, a fake nuclear alert and a $6bn heist: are we in the foothills of a hacking epidemic?As AI exposes flaws ...
📰 Hot fridges, a fake nuclear alert and a $6bn heist: are we in the foothills of a hacking epidemic?As AI exposes flaws in IT systems and geopolitics plays out on the cyber frontli...
Whenever I hear "we'll use Copilot for that document", I interpret that as "most of this document doesn't need to exist"...
Whenever I hear "we'll use Copilot for that document", I interpret that as "most of this document doesn't need to exist".Unfortunately, I hear it daily in the Corporate world now. ...