New York, NY, USA, October twenty first, 2026, CyberNewswire
Based on the 2026 Autonomous Brokers Menace Report, printed by Arrakis Safety, 76% of energetic brokers haven’t run in 60 days, however their credentials nonetheless work and their triggers nonetheless hearth. 44% haven’t run in 180 days. 24% of dormant brokers might be began by an unsigned webhook from outdoors the group.
The report relies on anonymized, unidentifiable manufacturing telemetry from 1000’s of AI brokers in Arrakis prospects’ environments, collected over a 12-week interval.
“There’s lots of dialogue about AI agent danger, and little or no knowledge”, stated Tal Baron, CEO and co-founder of Arrakis Safety. “When an worker leaves, their entry is eliminated the identical day. When an agent stops getting used, no person removes something, as a result of in most firms no person owns it”.
Different findings:
- 22% of brokers can ship content material to a mannequin run by a 3rd social gathering and act on the reply.
- 20% of brokers can move a process to a different agent with completely different permissions, and get again outcomes they aren’t allowed to provide themselves.
- 70% of brokers can attain delicate techniques, similar to cloud storage, AI companies, and advertising and marketing platforms, and virtually none of these connections have verified knowledge loss prevention (DLP).
- 82% of brokers carry not less than one publicity, and 46% of these carry multiple.
The complete report is out there at https://arrakis.safety/2026-threat-report.
About Arrakis Safety
Arrakis secures autonomous brokers by means of behavioral intelligence throughout three layers: what an agent is, what it does, and why. AI-SPM scans every agent’s construction, permissions, instruments, and metadata hourly, together with brokers that by no means run, surfacing shadow brokers, uncovered keys, and misconfigurations. AI-TDR displays actions in actual time for deviations from outlined guidelines and insurance policies. The behavioral intelligence engine fashions every agent from its personal historical past, posture, runtime, and context: the way it obtained there and what it might do subsequent. For every motion, it decides whether or not to proceed, escalate, or intervene.
Contact
Moranne Shechnick
Arrakis Safety
[email protected]
