AI agent security, in numbers.

of CISOs are confident they can authorize what individual AI agents can do. 46% can centrally control what agents access; 47% can identify all agents in their environment.

of organisations have no defined approach to limit AI agent access, or rely on predefined human access.

of security leaders are deeply concerned about excessive AI access not being properly reviewed. 21% govern agent access with shared credentials or broad-permission service accounts.

of companies say their AI agents have taken unintended actions. The most common: accessing unauthorized systems or resources (39%).

of CISOs say AI has access to core business systems. Only 16% say they govern that access effectively. 92% of organisations lack full visibility into AI identities.

of security professionals worry about the security implications of integrating AI agents.

of US CISOs say AI agents are a critical or significant security risk. 30% saw suspicious AI agent activity in 2025.

name over-permissioned access among the most significant pain points. 78% have no formally adopted policy for creating or removing AI identities.

of IT application leaders believe AI agents are a new attack vector. Only 13% strongly agree they have the right structures in place to manage them.

for a coding agent to delete a production database and its backups, using a token it found. The latest recoverable backup was about three months old, as the founder reported it.

of course data, and every automated snapshot, removed when an agent ran a destroy command against production. About 24 hours to recover.

attacker actions against Hugging Face over four days, by evaluation agents that broke out of their environment.

real systems ran a malicious package an evaluation agent published during roughly one hour online.

unsanctioned live-internet actions in 10 of 122 runs during cyber testing of AI agents.

of organisations that reported a breach of AI models or applications had no AI access controls in place. 13% reported such a breach. The global average cost of a data breach: $4.44 million.

of participants in a public prompt injection challenge got a generative AI bot to reveal information it was told to protect.

successful policy violations from 1.8 million prompt injection attacks on AI agents, including unauthorized data access and illicit financial actions.

of the target frontier models had at least one successful attack found against them in a 2026 indirect prompt injection competition.

forecast spend on securing AI in 2026, up 83%.

forecast for 2027, a further 68.7% rise.