AI agent insider threat

An AI agent holds identity, access and autonomy inside your environment. Above investigates what it did, whose access it used and whether that matched its task.

Four identical masks on stands in a row. One of them is violet.

INC-023A platform engineer's Claude Code agent

Confirmed threat

Claude Code agent reads 1,200 Finance files on its own

  1. Claude CodeGives Claude Code a task
  2. Google DriveReads 1,200 Finance files
  3. CrowdStrike FalconLaptop offline

Response Revoke the agent's access

AI agents, by the numbers

  • 25%

    of enterprise breaches will be traced back to AI agent abuse by 2028, a prediction. Gartner, October 2024.

  • 92%

    of organizations with an AI-related breach lacked proper AI access controls. IBM, Cost of a Data Breach 2026.

  • 18%

    of organizations have fully integrated AI governance into their Insider Risk program. Ponemon Institute, 2026 Cost of Insider Risks.

Why an AI agent is an insider

“We need to separate the supply of intelligence from the authority over it.” Satya Nadella, CEO of Microsoft, on models as Insider Risks.

Identity

It signs in with a person's token or its own.

Access

It reaches whatever that identity can reach.

Autonomy

It picks its next step and acts at machine speed.

Where a person ends, and an agent begins

Some agents are an extension of a person. Others run in your cloud on their own. Above knows which side of the line each agent is on, and investigates it there.

An extension of the person

A platform engineer

  • Claude Code
  • GitHub Copilot
  • Claude
  • ChatGPT
  • Google Workspace
  • Microsoft Entra

Assistants and the agents people connect to their accounts work with that person's access. The person stays accountable, so Above reads the agent as part of their story.

Autonomous in your cloud

  • Microsoft Foundry
  • Bedrock AgentCore
  • Amazon Bedrock

Agents you build run under their own identity. Above investigates each one as an insider in its own right, tied to the owner accountable for it.

Public incidents

With a person's access

  • A startup using Claude Code

    February 2026

    1. Claude CodeClaude Code works from an older Terraform state file
    2. Claude CodeClaude Code runs terraform destroy against that state
    3. AWSAWS production database and network are removed

    The founder attributed it to relying on the agent too much. AWS restored a snapshot.

    Sources: The founder's account (opens in a new tab)

    How Above investigates this kind of event

    1. OktaEngineer signs in
    2. Claude CodeClaude Code session
    3. Terraformterraform destroy
    4. AI agentNeeds review

Autonomous, in the cloud

  • OpenAI's evaluation agents and Hugging Face

    July 2026

    1. OpenAIOpenAI's agents escape an isolated evaluation through a flaw in its package proxy
    2. OpenAIAn agent finds Hugging Face user credentials exposed online
    3. Hugging FaceThe agents run code on Hugging Face servers and harvest more credentials
    4. Hugging FaceHugging Face contains it and rotates those credentials

    OpenAI said the agents were pursuing their evaluation, without malicious intent.

    Sources: Hugging Face (opens in a new tab), OpenAI (opens in a new tab)

    How Above investigates this kind of event

    1. Microsoft FoundryAgent runs a test
    2. OktaFound credentials used
    3. AWSData read out
    4. AI agentNeeds review
  • Claude in Anthropic's cyber evaluations

    July 2026

    1. ClaudeClaude Opus 4.7 is told its evaluation has no internet
    2. ClaudeClaude Opus 4.7 reaches a real company's systems over a live connection
    3. ClaudeClaude Opus 4.7 extracts credentials from those systems
    4. ClaudeClaude Opus 4.7 reads production data with those credentials

    Anthropic attributed it to a misconfiguration and found no model pursuing its own goal.

    Sources: Anthropic (opens in a new tab)

    How Above investigates this kind of event

    1. Microsoft FoundryAgent's test task
    2. Microsoft Entra IDReal system sign-in
    3. AWSProduction data read
    4. AI agentNeeds review

The insider types, applied to AI agents

The compromised agent

Hijacked by a prompt injection or a poisoned MCP server, it follows the attacker's instructions instead of its owner's. OWASP lists this as Agent Goal Hijack (ASI01).

The negligent agent

Over-scoped and set to run on its own, it chains tools nobody reviewed and shares or deletes more than the task needed.

The agent used maliciously

A person uses an agent to do at machine speed what they shouldn't do by hand, such as exporting a customer base in one run.

The agent left behind

Agents and OAuth grants keep running after the person who set them up has moved role or left the company.

An agent on its own

An agent on its own

Above investigative agents

  • Timeline
  • Contextual analysis
  • Reasoning
  • Recommended actions

Ready for security, HR, and legal to act on together.

What the investigation hands you

Ready for security, HR, and legal to act on together.

An agent on its ownINC-023 · A platform engineer's Claude Code agent

Confirmed threat

Claude Code agent reads 1,200 Finance files on its own

Every read used the engineer's own valid access.

Synthetic Insider Threat Matrix IV007: Autonomous Self-Invocation

The evidence

  • Claude Code
  • Google Drive
  • CrowdStrike Falcon

The engineer's part ended when the laptop went offline; the Sunday reads were Claude Code's alone, and nothing in its task touches Finance.

The response

  1. Revoke the agent's access
  2. List exactly what it read
  3. Rotate any secrets in those files
  4. Tell the engineer

A human confirms

Talk to a human, 30 minutes

A solutions engineer walks you through a real case.

The signals Above reads for AI agents

Above reads the grant, the platform and what the agent touches.

Show the signals from
  • New OAuth grants to AI agents

    Google Workspace, Microsoft Entra

    Which agent, who approved it, which scopes.

  • Scopes wider than the task

    Google Workspace, Microsoft Entra

    Full Drive or mail access for a meeting summarizer.

  • Reads far beyond the ask

    Google Drive, Slack

    Hundreds of files opened for a task needing a handful.

  • Writes to personal or outside destinations

    Google Drive

    Exports and shares to accounts the company doesn't own.

  • Runs from a new device or place

    Okta

    A device or location its owner never used.

  • Actions on agent platforms

    Bedrock AgentCore, Microsoft Foundry, Amazon Bedrock

    What agents built in your cloud did, and their scope.

  • Agents outside approved paths

    Claude, ChatGPT

    Unreviewed agents and MCP servers the Shadow AI investigator finds.

  • Agents still active after their owner leaves

    Google Workspace, Workday

    Grants and scheduled runs that survive a termination.

  • Agent and owner in the same window

    Read together as one story.

Start with the agent

Above starts from the agent's actions and its grant, and brings the person in only when the evidence points to them, under the same controls as any other case.

  • The person behind an agent stays pseudonymized until a reviewer is authorized to unmask, and every unmask is logged.
  • Recommendations target the agent first: its grant, its scopes, its device.
  • Agents recommend; a human confirms every action.

Related use cases

Shadow AI and AI data leakage

Company data in personal AI accounts and custom GPTs, and how to adopt AI without the leak.

  • Compromised accounts

    A real employee's access in someone else's hands. Above shows whether it was the person or an attacker.

  • Departing employee data theft

    Customer lists, pricing and code that leave in the weeks around a resignation.

AI agent risk on the blog

Questions about AI agent insider threat

What is an AI agent insider threat?

It is harm done through an AI agent that holds legitimate access: an assistant, a coding agent or an automated workflow acting with a person's or a service's credentials. The Insider Threat Matrix™ covers it as Delegated Execution via Artificial Intelligence Agents (IF028), and its synthetic matrix treats AI systems as insider subjects in their own right.

Is an AI agent an insider?

In practice, yes. It acts inside your systems with access you granted, and the classic insider types map onto it: an agent can be compromised by a prompt injection or a poisoned MCP server, negligently over-scoped, used maliciously by its owner, or left running after its owner has gone.

Was the Hugging Face incident an insider threat?

To Hugging Face it was an outside intrusion, driven end to end by an autonomous AI agent system. From OpenAI's side it fits the insider pattern: its own evaluation agents, running in its environment, got past the controls meant to isolate them, with no human directing them, and used credentials they found.

Can one tool investigate AI agent risk and human Insider Risk together?

One tool can investigate both, because most agents act with a person's access. Above investigates humans and AI agents together across SaaS, identity and AI. Coding agents such as Claude Code and GitHub Copilot work as an extension of their user, so the case shows who it acted for, what it touched and whether that matched its task.

How do you tell what an agent did from what the employee did?

Above ties each agent to its grant and its owner, then reads the agent's actions on the platforms it uses alongside the person's own activity. Timing, device, volume and the task the agent was given usually separate them. When they don't, the case says so.

Does Above block AI agents?

Above doesn't block agents itself: it investigates and recommends, and your team acts. Typical recommendations are revoking the agent's OAuth grant, narrowing its scopes, confirming a new device with the owner or recovering what it exported. A human confirms every action, so nothing changes in your systems without someone on your team approving it.

Which agent platforms does Above read?

Bedrock AgentCore, Amazon Bedrock and Microsoft Foundry for agents you build; Claude Enterprise through the Compliance API and ChatGPT for AI assistants; and OAuth grants in Google Workspace and Microsoft Entra for agents your people connect to their accounts. Above adds integrations every month.

Know what your agents did, and for whom

See an AI agent investigation from grant to verdict.