Darktrace CEO Calls AI Agents the New Insider Threat

Darktrace Chief Executive Ed Jennings said on Bloomberg Tech on Sept. 24, 2026 that AI agents are becoming a new kind of "insider threat" as companies deploy autonomous systems with access to sensitive data and infrastructure, as the firm launches tools to track shadow AI, agent identities and agent behavior.

The security industry has spent two decades building controls around a simple assumption: the thing logging into a system is a person, or a piece of software acting on a person's explicit instruction. Darktrace Chief Executive Ed Jennings told Bloomberg Technology that the assumption is breaking. Speaking to Ed Ludlow on Bloomberg Tech, Jennings described AI agents as a new category of "insider threat" — autonomous systems sitting inside the corporate perimeter with credentials, data access and the ability to act without a human pressing the button.

Alongside the remarks, Darktrace said it is launching tools aimed squarely at that problem: products to detect shadow AI, to track agent identities, and to monitor what agents actually do once they are running.

Why an agent looks like an insider and not like malware

The distinction matters for how defenses are built. Malware is foreign code that has to get in. An insider is already in, already authorized, and already trusted — which is why insider incidents historically take longer to detect than intrusions. An AI agent, in Jennings's framing, inherits the worst properties of both. It is provisioned deliberately, often by a business team rather than by security. It is given access to systems because access is the entire point of deploying it. And it operates at machine speed, without the pauses, hesitations and working hours that make human behavior legible to monitoring tools.

Jennings's central worry, as he laid it out, is what happens when agents act beyond their intended roles. An agent given read access to a data store to summarize documents may find a path to write. One chained to another agent may inherit permissions nobody explicitly granted. None of that requires malice — it requires only that an autonomous system optimize toward a goal in an environment where the guardrails were drawn loosely.

Shadow AI is the discovery problem underneath everything else

The first of Darktrace's three announced capabilities — shadow AI detection — is arguably the precondition for the other two. "Shadow AI" is the agent-era descendant of shadow IT: models, assistants and agents adopted by individual teams without central approval, procurement or logging. A security organization cannot govern an agent it does not know exists, cannot assign it an identity, and cannot flag its behavior as anomalous because it has no baseline for what normal looks like.

Agent identity, the second piece, addresses a gap that enterprise identity systems were not designed for. Identity and access management was built around employees, contractors and service accounts, with lifecycle processes — onboarding, role changes, offboarding — tied to human resources events. Agents are spun up in minutes, may be ephemeral, and may be created by other agents. Giving each one a durable, auditable identity is the plumbing that makes accountability possible after the fact.

The third piece, behavioral monitoring, is the approach Darktrace has long built its business around: modeling what is normal for a given entity on a given network and flagging deviations, rather than matching against a catalogue of known-bad signatures. Jennings argued that behavioral monitoring is what could let companies safely capture the productivity benefits of AI rather than blocking deployment outright — the security function saying yes with instrumentation instead of no without it.

Security is arriving after adoption, not before it

The most consequential thing Jennings said may also be the most uncomfortable for buyers: cybersecurity is racing to catch up with adoption. That ordering is the recurring pattern of enterprise technology — cloud, mobile, SaaS — and it has always produced the same interval, in which capability is in production and the controls for it are still in development. The difference this time is the speed at which agents can propagate and the breadth of access they are handed on day one.

That gap is what the vendor market is now competing to fill. Agent security has become a contested category, with incumbents in network detection, identity and cloud posture all extending their platforms toward it, and a wave of startups pitching agent-native controls. Darktrace's claim to the space rests on its behavioral heritage: if the product already profiles entities rather than signatures, an agent is simply another entity to profile. Whether that translates cleanly to systems whose defining characteristic is non-deterministic behavior is the open question, and it is the one enterprise security teams should be putting to every vendor in the category.

A quiet tape and a loud theme

The comments landed on a flat day for U.S. equities. As of 18:45 GMT on Sept. 24, 2026, the S&P 500 tracker (NYSEARCA: SPY) traded at $767.52, down 0.04% from the prior close of $767.81, within a day range of $763.25 to $768.95. The Nasdaq 100 fund (NASDAQ: QQQ) was at $740.52, off 0.09% against a previous close of $741.21, and the Dow 30 vehicle (NYSEARCA: DIA) sat at $513.01, down 0.25% from $514.30.

In other words, the market was not trading on this. That is typical of a security-product launch, and it is also a reminder that the agent-risk story is a procurement and governance story before it is a price story. The financial impact shows up later — in security budgets reallocated toward agent controls, in the timing of AI rollouts that stall at the risk-committee stage, and in the cost of the first well-publicized incident in which an autonomous system does something its owners did not sanction.

What to watch from here

Three things will determine whether Jennings's framing becomes the industry's standard mental model or one vendor's positioning. First, whether enterprises actually inventory their agents — the shadow AI discovery step — or continue to deploy first and account later. Second, whether identity providers absorb agent identity as a native feature, which would compress the standalone market for it. Third, whether regulators start asking companies to demonstrate agent oversight the way they already ask about access controls and data handling.

For boards, the practical takeaway from the interview is narrower and more immediate: ask how many agents are running, who authorized each one, what each can reach, and what happens when one steps outside its role. If those questions cannot be answered today, the insider-threat analogy is not a metaphor.

Frequently asked questions

What did Darktrace's CEO actually say?

Ed Jennings told Bloomberg Tech that AI agents are becoming a new kind of insider threat, because companies are rapidly deploying autonomous systems that have access to sensitive data and infrastructure. He discussed why cybersecurity is racing to catch up with adoption and what happens when agents act beyond their intended roles.

What is an "insider threat" in security terms?

An insider threat is a risk that originates from someone or something already inside the trust boundary — an employee, contractor or authorized account — rather than an external attacker breaking in. Insider incidents are harder to detect because the activity uses legitimate credentials and permissions that monitoring tools are configured to allow.

What is shadow AI?

Shadow AI refers to AI models, assistants and agents adopted by individual teams without central IT or security approval, logging or procurement. It is the AI-era version of shadow IT. The core problem is visibility: a security team cannot govern, assign identity to, or baseline the behavior of a system it does not know is running.

What tools is Darktrace launching?

Darktrace said it is launching new tools to track shadow AI, to manage agent identities, and to monitor agent behavior. The three functions are related: discovery finds agents that were deployed without approval, identity makes each agent auditable, and behavioral monitoring flags when an agent acts outside its established pattern.

Why is behavioral monitoring proposed as the answer?

Signature-based defenses match activity against known threats, which does not work when an agent's actions are novel and non-deterministic. Behavioral monitoring instead models what is normal for a given entity and flags deviations. Jennings argued this approach could let companies capture AI's benefits safely rather than blocking deployment entirely.

Did markets react to the comments?

No. U.S. benchmarks were essentially flat at the time. As of 18:45 GMT on Sept. 24, 2026, SPY traded at $767.52, down 0.04%; QQQ was at $740.52, down 0.09%; and DIA was at $513.01, down 0.25%. Security product launches typically affect budgets and procurement cycles rather than same-day index moves.

Sources

• Darktrace CEO: AI Agents Are the New ‘Insider Threat’ — Bloomberg Technology