← All submissionsReturn Home ↗
@unix_jim's runlevel-1 Dojo

CyberSecurity: When Humans Become the Bottleneck

Agentic AI, bounded autonomy, and the future of threat hunting

by @unix_jim - (updated September 2026)
Bounded defensive autonomyAn independent policy gate separates AI investigation from constrained containment. Human authority defines policy and approves consequential remediation.InvestigatePolicy gateContainEvidence · HypothesesScope · Permission · LimitsBounded · AuditableHuman authority defines policy and approves consequential remediation.
30 second summary (Simona’s Version)

The important question is not whether humans remain “in the loop.” It is where human authority belongs when machine-speed components can operate faster than human deliberation and decision making.

Cybersecurity has spent decades putting humans into decision loops for good reasons.

A machine may identify an anomalous authentication event, suspicious process tree, or unexpected network connection, but an analyst is expected to determine what it means before consequential action occurs.

That model assumes something rarely stated explicitly: the defender has enough time to think.

Increasingly, that assumption now deserves closer examination.

Which defensive decisions still require human judgement, which can safely be delegated, and how much delay can security systems tolerate while making that distinction?

That is the real problem behind agentic cybersecurity.

The wrong argument: humans versus AI

Discussion of autonomous cybersecurity often collapses into two camps. One imagines AI agents defending infrastructure at machine speed. The other responds, correctly - that contemporary AI systems hallucinate, misunderstand context, inherit poisoned data, make brittle decisions, and can themselves become attack surfaces.

Both positions miss the more useful systems question:

Which decisions should occur at which layer of the control system?

A modern Security Operations Centre already contains substantial automation: correlation rules, endpoint response, identity-risk scoring, behavioural analytics etc. Hell SOAR workflows routinely make or recommend decisions without an analyst manually examining every underlying event.

Agentic AI changes the problem because the system can potentially select actions, gather additional evidence, invoke tools, revise a hypothesis, and determine subsequent steps from evolving context.

That increased flexibility is also where increased risk appears.

Threat hunting has a latency problem

Traditional threat hunting is often described as proactive security. Operationally however, much of it remains dependent upon human investigation:

Every transition consumes time. Machines are exceptionally good at searching large telemetry sets, correlating identifiers, retrieving context, and repeatedly applying deterministic procedures. We remain stronger at organisational context, ambiguous intent, novel business consequences, and deciding whether an unusual event is actually unacceptable.

The design error is to automate the entire chain merely because portions of it can be automated. The opposite error is equally real: requiring human approval for every intermediate operation can transform oversight into latency.

Behaviour matters more than labels

Hashes change. Infrastructure changes. Malware families mutate. Identities, workloads, and systems nevertheless have operational constraints.

Behavioural analysis is therefore attractive because it asks whether observed activity is consistent with expected operational patterns. But a baseline is not truth; it is a statistical representation of previous observations.

An attacker who gradually alters behaviour may contaminate the very history from which future behaviour is judged. An agent operating on compromised telemetry can consequently become confidently wrong.

A credible agentic security system therefore requires more than “an AI watching the network.” It requires telemetry provenance, confidence scoring, temporal comparison, segmentation, and mechanisms capable of identifying when the assumptions feeding an agent have changed.

The evidence argues against unrestricted autonomy

Microsoft Research’s ExCyTIn-Bench, published at ICML 2026, evaluates LLM agents on cyber-threat investigation using a controlled Azure environment spanning 57 security-log tables and 7,542 generated questions. The best reported model achieved a reward of 0.606, leaving substantial headroom.

That is not evidence that autonomous cyber investigation is useless. It is evidence that it is already capable enough to matter, while remaining imperfect enough to make unrestricted authority dangerous.

ENGINEERING INTERPRETATION

Capability without complete reliability is precisely the condition under which authority boundaries matter most.

Containment may be the critical boundary

Containment and remediation are NOT the same operation.

Containment

  • isolate a suspicious endpoint;
  • suspend a session;
  • revoke a token;
  • temporarily disable a credential;
  • restrict network connectivity;
  • quarantine a process;
  • preserve evidence.

Remediation

  • delete resources;
  • rebuild machines;
  • modify production infrastructure;
  • permanently change permissions;
  • remove accounts;
  • alter business processes.

The first category contains many actions that can be made bounded, reversible, and observable. The second contains actions whose consequences may be difficult or impossible for an AI system to predict.

AI authority should increase as action reversibility increases and consequence decreases.

NIST’s 2026 Cyber AI work similarly frames AI-enabled cyber defence as an opportunity that must coexist with governance, evaluation, attack-surface management, and explicit risk controls.

A five-level defensive autonomy model

0

Observe

Read telemetry. No environmental changes.

1

Investigate

Correlate evidence, query systems, construct hypotheses.

2

Recommend

Propose response actions; require human authorization.

3

Contain

Execute predefined, reversible actions within policy boundaries.

4

Remediate

Perform consequential infrastructure or business changes.

The numbering is less important than the principle: authority should be graduated rather than binary.

The agent becomes part of the attack surface

An autonomous security agent is software with identity, permissions, memory, tools, and communication channels. It should therefore be treated as a privileged security principal, not merely an analytical application.

That implies familiar controls: least privilege, credential isolation, strong identity, tool allow-lists, segmentation, auditable actions, immutable logging, rate limits, rollback, and independent monitoring.

The agent that decides an action should not necessarily be the component that authorises that action.

Cybersecurity spent decades learning not to give humans unlimited privileges. There is little justification for forgetting that lesson when the principal happens to be artificial.

We (humans), are not disappearing

Calling humans a 'bottleneck' produces the wrong conclusions; human judgement still remains essential. What deserves reconsideration is where we place it.

LESS

Inspect every alert.

MORE

Design the system that decides which alerts matter.

LESS

Execute every containment action.

MORE

Define which actions machines are permitted to execute.

LESS

Remain permanently inside every decision loop.

MORE

Determine the boundaries and escalation rules of those loops.

The question we should actually be asking

The debate surrounding autonomous cyber defence is frequently expressed as one of trust: Can we trust AI?

Security engineering has never depended upon absolute trust. We design systems assuming components can fail.

Agentic cybersecurity should be treated no differently.

The objective is not to construct an AI sufficiently intelligent that humans can finally trust it. The objective is to construct an architecture in which an imperfect AI can operate without its inevitable errors becoming catastrophic.

That means bounded authority, reversible actions, independent verification, continuous telemetry, explicit confidence thresholds, separation of decision and execution, and humans positioned where judgement matters more than reaction time.

The future of threat hunting may indeed involve humans stepping out of some operational loops.

The important question is which loops they leave, who defines the boundaries, and what happens when the machine is wrong.

References

  1. Lazer, S. J., Aryal, K., Gupta, M., & Bertino, E. (2026). A Survey of Agentic AI and Cybersecurity: Challenges, Opportunities and Use-case Prototypes.
  2. Wu, Y. et al. (2026). ExCyTIn-Bench: Evaluating LLM Agents on Cyber Threat Investigation. Proceedings of the 43rd International Conference on Machine Learning, PMLR 306.
  3. NIST (2026). NIST IR 8607 — Workshop Summary Report for “Cyber AI Profile” Hybrid Workshop #2.
  4. NIST NCCoE (2026). Cyber AI Profile Virtual Working Session Series: Extending the Technical Content.

Reference set checked against live publisher and institutional pages on 6 October 2026.

Additional evidence & attribution

External research and institutional documentation supporting the specific new arguments above. Interpretive claims and proposed operational frameworks are the author’s synthesis.