Geordie AI Ltd

11/18/2026 | Press release | Distributed by Public on 11/18/2026 04:02

Risk Intelligence: Finding the agent risks that only show up in context

Most security teams can now list the agents running in their organization. The harder part is knowing which of them put the business at risk. Geordie observes billions of agent activities and detects hundreds of thousands of risks each month to ensure businesses keep their AI agents operating as intended. A single agent can carry dozens of signals, from the tools it's configured with to the commands it runs frequently. Most are routine tasks, and the ones that matter often look under the radar too until they're read against what the agent can reach. Earlier in this series, we covered the record of what every agent does and how Beam acts on what it finds. This post covers one of the most business-critical steps in between.

A large consumer marketplace we work with saw this firsthand. Geordie flagged secrets stored in plain text in the configuration of the MCP servers its agents connect to, where any agent reading the file could pick them up. On one agent, that's a single finding. Spread across tens of agents and owners, it rose near to the top of the team's list.

Risk Intelligence was built to judge each risk from the whole agent, reading how it's configured and how it behaves as one picture. It ranks each risk by the harm it could do and the number of agents it reaches, and traces it to the tool and credential behind it. The security team knows where to start and where the fix goes.

What: Risk judged from how an agent is set up and what it does

Risk Intelligence closes the gap between what an agent is set up to do and what it actually does, by reading two things together:

  • Exposure findings: what could happen, based on how the agent is set up, such as an agent built with its creator's own credentials, which lets anyone who uses it to reach the creator's files.
  • Activity risks: what did happen, based on the actions the agent took, such as a credential passed in a tool call or a package installed from an untrusted source.
Exposure risks from how an agent is set up and activity risks from what it does, ranked together in one list.

Each risk carries a severity from Critical to Low. Findings that are accurate but harmless, such as an internal email address, are marked Informational so they stay visible without crowding the queue. Findings come from a library of scenarios that Geordie's AI risk research team updates as new threats are disclosed, covering exposed credentials, over-broad permissions, prompt injection, poisoned packages and data leaving through a skill.

When a trojanized Claude Code extension appeared in Cursor's marketplace on 29 October, Geordie added detections for the exposed agents within hours, with step-by-step guidance for the SOC on preserving evidence and cleaning up each machine. Teams can also add their own scenarios, with Ask Geordie drafting them for review.

How: From hundreds of signals to the few that matter

Context is what turns a signal into insight. Geordie reads each agent's configuration, the tools it connects to and its activity as one picture, which lets Risk Intelligence do a few things that matter to the team acting on it:

  • Risk in the context of the whole agent. A finding reflects what the agent can reach, what it did with that access and where it sits in the business, so an agent querying a database reads differently in engineering than in marketing.
  • Deep enough to act on. Each risk names the tool, user, credential and permission it depends on, so the fix goes to the right place.
A Claude Code agent's risks, each tied to the tool it depends on.
  • Wide enough to trust. Risks cover shadow and unsanctioned agents and the skills, extensions and plugins that tools watching a single surface never see.
  • A path the team can follow. Each risk shows how it forms. The classic case is the lethal trifecta: untrusted input comes in through one tool, the agent can reach sensitive data, and another tool can send that data out.
The lethal trifecta, where three routine tools add up to one risk path.
  • A priority that reflects reach. Each risk is ranked by its severity and the number of agents it affects, so a hard-coded secret on multiple agents across various owners can rise above a more severe risk on a single, isolated agent.
  • A view for each team. Priorities group risks for the SOC, for AppSec and for configuration hygiene, and each agent owner sees the risks on their own agents.
Top priorities, where a hard-coded MCP secret rises above a Critical risk on a single agent as its reach grows across agents and owners.

Why: Attention goes where it's needed

The number of agents in most organizations is growing faster than the teams responsible for them. At one enterprise with hundreds of thousands of agents operating each day, Geordie prioritized every finding by its severity and the number of agents it reached. The few that put the business most at risk rose to the top, and each one was delivered into the workflow tools the security team already works in. That prioritization gives the team a defensible answer to what comes first, backed by what each agent can reach and what it did. Each finding can also flow into the risk register a GRC team already runs, along with its evidence, so the same view that guides the security team informs how the organization tracks risk.

Ranking well depends on context, because some of the most important findings look routine at first. At a financial services and investment firm, an agent was mining bitcoin while coordinating with an agent belonging to an external adviser, and Geordie surfaced the activity. At a fintech company, a finance leader had connected his agent to Telegram through a skill and was sending sensitive company information out to third parties.

Surfacing findings like these is the first step toward resolving them. Every finding carries the steps to fix it and reaches its owner through the workflows the team already uses. Where a finding calls for a control, Geordie's inline remediation engine Beam suggests a policy that can guide the agent, pause it for approval or stop it at the moment of risk.

Evaluation: Questions to ask about any tool that claims to assess agent risk

Before your next risk review, we recommend asking these questions of any vendor that assesses how risky your agents are:

  • Does it judge risk from what an agent is set up to reach as well as what it did, or only from the traffic or device events it sees?
  • Does the same action read differently depending on the agent's team and what it can reach, or is every match treated the same way?
  • Can it tie a risk to the specific agent and its owner, and name the tool, credential and permission involved?
  • Does it cover the skills, extensions and plugins that never send traffic through a gateway, and the agents that never run on a monitored laptop?
  • Can it show how a risk forms across an agent's tools, such as untrusted input reaching sensitive data with a way out, or does it flag each tool on its own?
  • When a new threat to agents is disclosed, how quickly does it show which of your agents are exposed, and with what steps to fix them?

To see how Geordie acts on these findings, read Beam: Guiding agents as they work, with tiered escalation and Beam in practice. Teams that want to see how their agent risk ranks can get in touch through geordie.ai.

Geordie AI Ltd published this content on November 18, 2026, and is solely responsible for the information contained herein. Distributed via Public Technologies (PUBT), unedited and unaltered, on November 18, 2026 at 10:02 UTC. If you believe the information included in the content is inaccurate or outdated and requires editing or removal, please contact us at [email protected]