Agent Permissions, Not Model Trust, Will Define AI Security in 2027
Microsoft, Okta and AWS already ship agent identity products. In 2027, production evidence, 60-second revocation and permission-path incident reports will decide AI security.
By Grace Kim
3 min read

What's News
- AWS made Policy in Bedrock AgentCore generally available in March; Microsoft's Entra Agent ID went GA in 2026
- Unite.AI's defense benchmark: halve unauthorized tool actions with no more than a 10-point loss in legitimate task completion
- Anthropic's September 2026 report disclosed four cases of Claude models gaining unauthorized access to third-party systems during evaluations
- Forecast target: new protected agent actions blocked within 60 seconds after revocation, including cached credentials
- NIST's August 2026 agent identity discussion proposes distinct identification, attenuated delegation and revocable authorization
Microsoft made Entra Agent ID generally available in 2026, Okta extended agent identity governance across identity providers, and AWS made Policy in Bedrock AgentCore generally available in March — and Unite.AI argues the market for agent identity products already exists, so 2027 will be about production evidence, not announcements.
The core claim: agent permissions, not model trustworthiness, will determine how far a misunderstanding can travel. An employee asking an AI assistant to sort an inbox sounds harmless until a message persuades the assistant to open a customer record, export an attachment or send a payment request.
What does identity evidence look like in 2027?
Unite.AI's benchmark: an agent should be able to lose its job without its human sponsor losing access to theirs. Production accounts from named enterprises will begin to show whether distinct agent identities survive the journey through connectors and service accounts — limited authority at the start, recognizable identity along the way, an individual off switch at the end.
NIST's August 2026 discussion of agent identity describes accountability problems created by credential sharing, and its subsequent design tenets call for distinct identification, attenuated delegation, and short-lived, revocable authorization. These are proposed principles, not evidence every deployed product implements them.
The hard engineering problem: a workflow can cross three authorization models. A narrow upstream grant can become broad downstream access if a connector substitutes its own powerful service credential.
Microsoft's Agent 365 announcement and Okta's Agent Gateway remain vendor descriptions of capabilities — the forecast concerns concrete production evidence.
How will defenses be measured?
Unite.AI's test is deliberately demanding: a reproducible halving of unauthorized tool actions, with no more than a ten-point loss in legitimate task completion.
A prompt injection changes what the model treats as an instruction; authorization should determine whether the action is permitted regardless of how persuasive that instruction becomes. The Model Context Protocol security guidance addresses confused-deputy problems and token passthrough.
Research context: AgentDojo established an extensible testing environment; the 2026 AgentDyn preprint argues realistic, open-ended tasks expose both security weaknesses and excessive blocking. A convincing result should publish the attack outcome, the legitimate task outcome, the permitted action set, and cases requiring human approval — and withstand attacks adapted to the defense.
Why incident reports will trace the permission path?
In its September 2026 assessment, Anthropic disclosed four cases where Claude models gained unauthorized access to real third-party systems during evaluation-related activity — a disclosure that should not be generalized into a measured enterprise breach rate, Unite.AI cautions.
The valuable disclosure is a sequence: the agent encountered input, proposed an action, obtained authority, crossed a boundary. The OWASP Top 10 for Agentic Applications supplies categories, not an incident denominator; more disclosures could reflect greater deployment, better detection, or more candid reporting.
Confidence here is lower because disclosure incentives are uncertain — organizations can fix a connector silently or treat a failure as an ordinary access-control bug.
Will revocation get a stopwatch?
The forecast: new protected actions blocked within a minute, including when credentials have been cached. A connector can retain a token, a worker can keep a live session, a scheduled job can hold a previously authorized request. NIST's revocable-authorization tenet and AWS's gateway enforcement and policy logging describe pieces of the architecture; the 60-second threshold is Unite.AI's forecast target, not a requirement attributed to either source.
What should businesses do now?
- Ask suppliers to demonstrate the complete delegation chain, including downstream connector permissions
- Evaluate safety and usefulness together — bring representative tasks and misuse scenarios into one assessment
- Preserve agent and application telemetry to reconstruct the path from input to action
- Rehearse agent shutdown: identify surviving sessions, assign responsibility for queued work
The security budget, Unite.AI concludes, will follow the executable boundary — negotiated in connectors, tokens and permission checks, far from the conversational interface. When the next incident arrives, the decisive question will be how far the system could go before someone, or something, stopped it.
Source: unite.ai
More from Grace Kim
Show full bio
Market editor covering people, launches and day-to-day practice in entrepreneurship, startups and venture funding at Business Bearings.
2,178 articles