#agentsecurity — Public Fediverse posts
Live and recent posts from across the Fediverse tagged #agentsecurity, aggregated by home.social.
-
82% of enterprises are running AI agents they don't know about.
That number came out of #RSAC Conference 2026 — and it wasn't the most alarming stat on the table.
Sean Martin sat back down with Itamar Apelblat, Co-Founder and CEO of Token Security, to unpack what he heard walking the show floor and what the CSA data now makes impossible to ignore: 65% of organizations have already had an AI agent-related incident in the last twelve months. 82% found agents in their environment that nobody authorized. Only 21% have any formal process to retire an agent when it's done.
Discovery alone is not governance. Intent-based enforcement is. That's where this conversation lands — and it's worth your time.
A huge thank you to the team at Token Security for joining Sean Martin and Marco Ciappelli on this journey — both on the floor at #RSAC2026 and in the recap. We loved sharing your story and we're looking forward to many more conversations ahead. 🙌
📍 Where are we headed next? Glad you asked: Infosecurity Europe and Black Hat USA — see you there.
🎙️ Recap: https://youtu.be/ZeI5bSbQ070
🎙️ On Location: https://youtu.be/uWjCQC3LnaY
🌐 RSAC Coverage: https://www.itspmagazine.com/rsac
🌐 Next Coverages: https://www.itspmagazine.com/technology-and-cybersecurity-conference-coverage#TokenSecurity #AIAgents #AgentSecurity #CyberSecurity #CISO #CloudSecurity #AIGovernance #IdentitySecurity #CSAReport #InfoSec #RSAC2026 #InfosecurityEurope #BlackHatUSA #CyberSecurityPodcast
-
@Em0nM4stodon AI agents automate this exact data escalation pattern. Start with query, accumulate session context, then cross-session memory, then behavioral patterns. Each layer justified as a 'performance improvement.'
EnergenAI LLC tracks this as Context Creep: automated data ratcheting at inference speed, no human approving each step.
VAULT audits what agents actually retain vs what they claim to need:
https://the-service.live/scrub?ref=mastodon-contextcreep -
Sears exposed customer chatbot conversations — call recordings and text chats — to anyone on the web (WIRED, March 2026).
No classic vulnerability exploited. The AI system ran outside its configured scope with nothing monitoring the behavioral layer.
Endpoint security tools protect processes on machines. AI agents introduce a different question: who monitors what the agent decides?