Shadow AI gets its own threat detector

On 27 October 2026, Exeon Analytics announced Exeon.AgentGuard, a new module that extends its behavioural threat detection into AI agent activity (Help Net Security, 27 October 2026). The pitch is specific: visibility into approved AI use, shadow AI nobody sanctioned, and what autonomous agents actually do once connected to internal tools and external services.

Exeon built its name on network detection and response, watching traffic for the behaviour that signals a breach rather than relying on signatures. AgentGuard applies the same logic one layer up. It runs on data sources the organisation already has, with no new agents to deploy, and looks for deviations from expected agent behaviour: unauthorised data transfers, privilege misuse, actions on critical systems that fall outside an agent's normal pattern. Gregor Erismann, Exeon's CEO, framed the goal plainly: the product exists to help organisations "understand what their AI agents are doing and detect unexpected behaviour" (Help Net Security, 27 October 2026).

Why this is not an isolated launch

Exeon is reacting to a gap that is already well measured elsewhere.

Reco's State of Agent Security 2026 (26 August 2026), built from anonymised enterprise telemetry and analysis of 500 public Model Context Protocol servers, found 80% of AI tools running with no oversight at all, and, in smaller organisations, 414 unsanctioned AI tools per 1,000 employees. Two years into enterprise AI policy work, the unmanaged estate is not shrinking.

LayerX Security's 2025 enterprise browser telemetry found that 77% of employees paste data into generative-AI tools, and that around four in five of those sessions run through unmanaged personal accounts, outside single sign-on and outside any data-loss control the employer owns (LayerX Security, 2025; coverage: The Register, 7 October 2025).

Both numbers describe the same underlying fact from different angles: AI use inside the enterprise is running ahead of any system built to see it, and that gap now extends past chat tools into agents that act on systems rather than just answering questions. A security vendor building a detection product for exactly that gap is a market responding to a measured condition, not inventing a problem.

What it means for a regulated enterprise

An agent that can reach HR, ticketing, payroll or a customer database does not announce when its behaviour drifts from what it was approved to do. There is no leaver process for an agent, no manager who notices it is acting strangely, and no access review that catches it, because access reviews enumerate people, not autonomous processes.

That absence is becoming a compliance gap rather than just an operational one. The EU AI Act's record-keeping obligations assume an organisation can state what a system did; a security team that only finds out about an agent's actions when a detection tool flags them after the fact is closer to incident response than to governance. NIS2 puts cybersecurity risk management under personal management accountability, and an undetected agent moving data to an external service sits squarely inside that obligation. A detection layer answers "did something unexpected happen", which is necessary. It does not answer "was this agent authorised to do that in the first place", which is a different question and a prior one.

What actually addresses it

Behavioural detection and identity governance solve different halves of the same problem, and an enterprise needs both, not one standing in for the other.

Detection, the mechanism Exeon is building, watches runtime behaviour against a baseline and flags divergence after the fact. It is valuable precisely because most organisations have no visibility at all today, so even retrospective alerting is an improvement on silence.

The complementary mechanism is upstream of behaviour: knowing, before an agent acts, which credentials it holds, what those credentials actually permit, and who is accountable for the agent in the first place. An agent with a narrowly scoped, revocable identity and a logged record of every tool call gives a detection system something concrete to compare against (approved intent versus observed action) rather than just a behavioural baseline with no ground truth underneath it. Without that upstream record, detection can tell you something unusual happened; it cannot tell you whether the agent was ever supposed to be able to do it.

What to check on Monday

Ask three questions this week, without buying anything.

List every AI agent with access to a production system or customer data, and name an owner for each one. If no single document can produce that list today, the organisation has agents it cannot currently govern, regardless of what detection tooling sits on top of them.

For each agent on that list, ask what credential it uses and what that credential permits, not what it was intended to permit. Service accounts cloned from another integration routinely carry more authority than anyone approved.

Ask how an unexpected agent action would reach a human today. If the honest answer is "it wouldn't, unless something broke downstream," that is the gap a vendor like Exeon is now selling a product to close, and it is worth treating as a gap even before any tool is purchased.

How AANCER answers

AANCER's role sits upstream of behavioural detection rather than in competition with it. Each Certified Agent carries a signed passport issued by your own per-install certificate authority, naming the tools, data and spend it may reach, so there is a concrete, scoped record of intended authority for any detection system, in-house or third-party, to compare observed behaviour against.

Every tool call an agent makes is written to an append-only, tamper-evident audit ledger, which gives a security team the "what did this agent actually do" record without waiting for a behavioural anomaly to surface it. Revocation takes effect in under 30 seconds, so once unexpected behaviour is flagged, cutting an agent's access does not require redeploying anything or taking down the systems it was talking to.

This does not replace network-level behavioural monitoring of the kind Exeon.AgentGuard provides. It gives that monitoring a ground truth to work from, which is the piece most agent deployments are currently missing entirely.

Sources

• https://www.helpnetsecurity.com/2026/10/27/exeon-expands-threat-detection-to-shadow-ai-and-autonomous-agents/
• Reco, State of Agent Security 2026, 26 August 2026
• LayerX Security enterprise browser telemetry, 2025 (via The Register, 7 October 2025)