Ai Agents Need Identities Too

The modern enterprise attack surface is not a network edge; it is a directory. Tens of thousands of human identities, and — increasingly — multiples more non-human ones: service accounts, API keys, CI runners, and now AI agents with standing credentials. Attackers noticed years ago that stealing a valid credential beats exploiting a vulnerability: no malware, no exploit signature, just a login that looks almost right. The majority of serious cloud incidents now begin with a compromised identity, not a compromised host.

Almost right is the operative phrase. A stolen credential logs in from a slightly wrong place, at a slightly wrong hour, and touches resources slightly outside its habit. Each signal alone is dismissible — which is exactly why rule-based detection drowns: the rules fire on travellers and contractors all day, and the SOC tunes them down until the real one sails through.

Catching identity abuse requires knowing each identity's normal — its hours, geographies, volumes, and peer group — and evaluating every anomaly against that baseline in context. That is a per-entity statistical problem multiplied by a hundred thousand entities, refreshed continuously. No human team maintains that. It has to be computed.

SIEM economics are the quiet scandal of enterprise security budgets. Per-gigabyte pricing means the bill scales with data growth, not with value delivered — and security telemetry grows 30 to 40 percent a year as estates move to cloud. Security leaders now spend real engineering time deciding which logs not to collect, which is a sentence that should alarm every board member: the team charged with seeing everything is financially incentivised to look at less.

The workarounds have become an industry of their own. Routing verbose sources to cold storage, sampling flow logs, dropping DNS queries, keeping ninety days when investigations routinely need a hundred and eighty. Every one of these decisions is invisible until an incident, at which point the responder discovers the exact telemetry needed to scope the breach was filtered out to save four thousand dollars a month. The fine for the breach starts at seven figures.

Meanwhile the SIEM itself does less than buyers assume. It stores, indexes, and alerts. The expensive part — deciding what an alert means — is left to the customer. Enterprises effectively pay a premium tax on raw material and then pay again, in analyst salaries, to refine it.

Enterprises that give up on in-house coverage buy MDR, and most discover the same frustration within a year: the service is a black box. Alerts go in, closed tickets come out, and the reasoning in between is invisible. When the quarterly report says '14,000 alerts triaged, 12 escalated', there is no way to verify that the 13,988 closures were sound. You are asked to trust, without evidence, the same category of overloaded human analyst you were trying to escape — just employed by someone else.

The contract structure makes it worse. Response actions require a phone call and a change window. Data lives in the provider's tenant, so leaving means losing history. And when an incident does slip through, the post-mortem devolves into a liability negotiation between your counsel and theirs about what 'detected' means in schedule 3 of the SLA.

None of this means managed outcomes are the wrong goal — 24/7 coverage without building a night shift is exactly what most enterprises need. The failure is opacity. A managed detection function is only trustworthy if every verdict it renders arrives with its evidence attached: the questions asked, the answers found, the weights applied. Transparency is not a nice-to-have in MDR; it is the entire difference between a service and a liability shield.

The average tenure of a tier-1 SOC analyst in a large enterprise is under two years, and exit interviews are monotonous in their consistency: the work is repetitive, the false-positive rate is demoralising, and the path to interesting work is blocked by the queue itself. Organisations respond with wellness programmes and retention bonuses, treating an architectural problem as an HR problem. No bonus makes it fulfilling to close the same impossible-travel false positive forty times a week.

The cost compounds quietly. Every departure takes environment-specific knowledge that never made it into a runbook — which service accounts behave strangely at month-end, which subnet the scanner lives on, which VP travels constantly and trips geo-velocity rules. The replacement starts from zero, triages worse for six months, escalates more noise to tier 2, and the seniors who should be hunting spend their days re-answering questions the departed analyst had already answered.

This is the hidden tax of the human-queue model: the institution never accumulates judgment. Knowledge lives in people, people leave, and the SOC's effective experience level stays permanently junior no matter how long the SOC has existed. Any serious fix has to move that judgment into a system that does not resign.

Every enterprise security team now serves two masters: the attackers and the auditors. SOC 2, ISO 27001, PCI DSS, DORA, HIPAA — each framework wants proof that monitoring exists, that alerts are investigated, that incidents follow procedure, that access reviews happen. Assembling that proof is a quarterly scramble of screenshots, CSV exports, and reconstructed timelines, performed by the same senior people who are supposed to be hunting threats.

The dirty secret is that most of this evidence is theatre. A screenshot of a dashboard proves the dashboard existed on the day of the screenshot. It says nothing about whether the alert at 3 a.m. on a Sunday in February was actually worked, by whom, and on what basis. Auditors accept it because nothing better is usually available, and everyone in the room quietly knows it.

Regulators are tightening. DORA and the SEC disclosure rules ask for operational reality, not intent — how fast incidents were detected, classified, reported. That standard cannot be met retroactively with screenshots. It requires an operating model where every detection, verdict, and response action generates its own tamper-evident record as a by-product of doing the work, not as a separate documentation exercise.

Every enterprise has automation it is afraid to turn on. The SOAR playbook that could disable a compromised account runs in 'create a ticket' mode, because the one time it ran for real it locked out a departmental VP during quarter close. The lesson organisations draw — automation is dangerous — is the wrong lesson. The right one is that automation without graduated governance is dangerous.

The change board's questions are legitimate: What exactly will this action touch? What happens if the verdict is wrong? Who approved this scope, and how do we undo it? Most automation platforms answer none of these. They offer a run button and a log, which is why enterprise response automation stalls at sending notifications while actual containment still waits for a human with production access and a change ticket.

Closing the gap requires treating autonomy as something earned per action class, not granted globally: recommend-only until precision is proven, approval gates where blast radius warrants, dry-run previews before anything writes, rollback state recorded for everything reversible, and a kill switch that stops it all instantly. With those primitives, automation stops being a leap of faith and becomes a controlled delegation — which is the only form of delegation an enterprise should accept.