AI Agents' Stolen Keys Unleash Security Chaos

A September 2026 survey of 102 enterprises highlighted critical security gaps in managing AI agent access. Many organizations lack efficient methods to revoke access, with 17 reporting that some or most of their agents share credentials. This sharing significantly complicates security, as revoking one agent's access could disrupt others.

Recent incidents underscore these risks. In early September 2026, a vulnerability in AWS Bedrock AgentCore allowed temporary AWS credentials to be exposed via a single prompt. Due to excessive permissions, these credentials could control all agents in an account. Similarly, Fireworks AI disclosed a breach where compromised internal credentials accessed customer data, suspected to be API keys and tokens.

Regarding revocation speed, only 14 of the surveyed enterprises could revoke access within minutes. Forty-five could do so within hours, while 17 estimated needing over a day or a credential trace. Ten companies were unsure if they could revoke all agents at all, and another 14 did not know how their agents authenticate.

Experts emphasize that AI agents fundamentally differ from conventional service accounts, making revocation more complex. The speed at which agents operate can outpace credential rotation cadences. Furthermore, shared credentials make it difficult to pinpoint which agent is making a request, increasing the risk of broader system impact during revocation.

NIST and CISA's Interagency Report 8587, finalized on September 15, 2026, recommends access and identity tokens live no more than an hour. However, some systems, like AWS AgentCore's, used temporary credentials with overly broad permissions that lasted longer than recommended, despite expiring on schedule. This highlights a disconnect between credential lifecycles and the permissions they wield.

Many organizations struggle with asset management for their AI agents. Forty respondents reported a confirmed agent security incident or near-miss in the past year, yet 12 do not track agent incidents. Twenty could not specify the number of agent identities their organization uses, indicating a significant blind spot. Seventy-two companies plan to adopt new agent security solutions within the next 12 months, but success hinges on addressing credential sharing.