Artificial Intelligence
Article urges discovery-first AI agent security
The article said organisations should inventory AI agents before enforcing policy, citing Veeam survey figures, a METR intrusion example and SANS guidance that placed visibility ahead of control.
- The article said Veeam found that 70% of organisations already had AI workflows touching sensitive corporate data without full oversight.
- It said 67% reported that IT could not fully track the autonomous workflows employees were building.
- The article described a METR example in which an attacker found a personal EC2 instance, bypassed authentication and obtained a model-provider API key.
- It said no single monitoring source could see all agent activity across the network, endpoint, browser and SaaS layers, and it ended by promoting SANS materials and a December course.
The article said AI agent security had to begin with discovery rather than enforcement. It cited Veeam figures about untracked workflows and said the SANS checklist put inventory before controls.
It treated agent use as a form of shadow IT and warned that blocking tools before mapping what existed could disrupt legitimate use. As a recent example, it described an intrusion at METR in which an attacker found a personal EC2 instance, bypassed authentication and took a model-provider API key.
The article said that key was then used for three weeks, consuming the equivalent of $600,000 in tokens, while METR’s dashboard did not show rate-limited requests. It also said one monitoring layer was not enough, because activity could sit in the network, endpoint, browser or SaaS layers.
It said organisations needed correlated telemetry, separate agent identities and continuous review, and it named Douglas McKee when arguing that agent tool access should be treated as its own identity and policy problem. The article also said a kill switch only mattered once the agent could be identified, then closed by promoting SANS training for December 2026.
Named in this story
People
- Douglas McKee
- was named in advice about agent identity and policy
Companies
- Veeam
- provided the survey figures on AI workflow visibility
- Hugging Face
- was referenced in the earlier intrusion evaluation
- OpenAI
- was named in the evaluation of its agents
- AWS
- was one of the cloud platforms used as a visibility example
Organisations
- SANS
- published the checklist and promoted the training
- METR
- appeared in the intrusion example involving an AI agent
Products and systems
- Zero Trust for AI Agents: The Security Checklist
- was the checklist the article used as its framework
- EC2
- was the cloud instance the attacker found
- Azure
- was the other cloud platform used as a visibility example
- LiteLLM
- was cited as an example of an LLM gateway
- SEC530: Defensible Security Architecture and Engineering, Implementing Zero Trust for the Hybrid Enterprise
- was the course promoted at the end
How the source tells it
The article read as a technical advisory with a promotional SANS finish, using urgency more than overt alarm.
- urgency repeated imperatives and checklist-style sequencing
- vendor boosterism a closing push for SANS guidance and training