Threat Intelligence

Detecting Agentic AI Attacks: Signals, Hunts, and Queries

Photo: Markus Spiske (CC0 1.0)
In this article5 sections

Why agentic attacks look different in logs

Detecting agentic AI attacks means hunting machine-speed behavior rather than a single signature: an autonomous framework that probes, fails, retries a variation of the same technique, and escalates — all inside a session a human operator would need hours to perform. Your first two minutes of triage should score two dimensions at once: request rate and technique diversity, per identity, per session.

The FBI’s FLASH alert of 13 November 2026 describes the ARTEX framework as able to automate “multiple phases of cyber intrusions at machine speed,” and notes that some intrusions reached privileged access in under three hours. Microsoft’s behavior-based guidance points at the same telemetry shape: an agentic framework sends a wide variety of requests in a brief time, retries variations on the same technique, or jumps to the next stage at machine speed.

That changes what suspicious looks like in three concrete ways:

  • Volume without depth. Hundreds of API calls across dozens of distinct actions in minutes — reconnaissance a human tester would spread across a full day.
  • Near-duplicate retries. The same call with one parameter changed, failing the same way, ten times in a row. People rarely retry identically; agents do constantly.
  • Clustered errors. AccessDenied, NoSuchKey, 403 and 404 responses arriving in bursts as the agent maps the edges of your permission model.

Build the baseline before you build the alert: for your top cloud identities, measure distinct CloudTrail eventName values per ten-minute window. Flat and low is normal. A sudden ramp in distinct event names from a service account that normally performs two or three actions is the cheapest agentic-attack heuristic available without buying anything. Pair it with a hard look at why that identity can do those actions at all — a Zero Trust Architecture review of agent and workload entitlements usually removes half the attack surface an automated probe is looking for.

Identity and cloud control-plane signals

CloudTrail is where most agentic campaigns surface first, because the control plane records every probe even when the action fails. Datadog Security Labs observed ARTEX in AWS environments in October 2026 and published patterns worth turning into rules today.

Anonymous S3 access at session start. Sessions often open with unauthenticated S3 requests: userIdentity.type of AWSAccount, accountId of “anonymous”, and GetObject or ListObjects calls. Ordinary workloads essentially never call S3 anonymously.

Long-lived keys used for log mining. With stolen long-term access keys, the most frequent action Datadog observed was CloudWatch FilterLogEvents, always with limit 1000 — an agent mining your logs for secrets. Search terms included ?mongo, ?172., ?10., ?password and ?AKIA.

Cognito sign-up abuse. Another observed initial-access path was anonymous sign-up against public Cognito user pools: SignUp, InitiateAuth, RespondToAuthChallenge, GetUser and GetTokensFromRefreshToken, then short-term credentials under a CognitoIdentityCredentials role session, then read-only discovery: ListBuckets, ListSecrets, DescribeParameters, DescribeLogGroups, GetRestApis, DescribeRepositories and ListFunctions20150331.

Vended credentials used off-cloud. Credentials issued to an EC2 instance role being used from a non-AWS IP is among the highest-signal alerts you can write (Datadog). The same logic applies to any role session appearing from outside your cloud provider’s ranges and corporate egress.

Application-layer abuse. Microsoft observed Storm-3358 downloading the WISP, vshell and AdaptixC2 RATs with netcat and curl, and using anonymous file upload and server-side request forgery against internet-facing applications. Your WAF, proxy and application logs carry the first evidence in those cases, not CloudTrail.

Write these as narrow rules with owners: any anonymous S3 principal; FilterLogEvents with limit 1000 from an IAM user holding long-lived keys; more than a handful of AccessDenied events per identity per hour; a Cognito SignUp spike from a single ASN; and any role-session activity from an IP outside your approved ranges.

Endpoint and network signals

Agentic frameworks use what is already installed. Microsoft notes their default tools — Nmap, curl, wget, Python HTTP clients, and headless Chromium via Playwright — may leave default user agent strings, which makes proxy, WAF and web server logs searchable without deep packet inspection.

On endpoints the signal is process lineage and burst behavior, not known-bad binaries: an application server spawning curl or wget, one parent launching dozens of short-lived children in a minute, or ncat-style tooling appearing where it has no business case. Sysmon Event ID 1 (process creation) and Windows Security Event 4688 give you parent-child pairs; on Linux, auditd execve records or eBPF process telemetry do the same. Alert on interpreter-and-downloader chains, then review the command lines — an agent’s argument lists are unusually long and repetitive.

In network data, look for scanning and exfiltration footprints: a single internal host opening hundreds of short sequential connections to many destinations in VPC Flow Logs or firewall logs; DNS queries with high entropy or long hex and numeric labels from a host that normally resolves a handful of names; HTTP requests with default tool user agents. Datadog’s published indicators include the IPs 103.54.154.40, 154.64.252.77, 154.36.136.197 and 169.128.55.214, plus user agents artex-readonly/1.0, artex-pentest/1.0, artex-audit/1.0, Mozilla/5.0 ARTeX, artex-spawn/1, artex-verify/1.0 and ARTEX-worker/1.0. Exact strings are ephemeral; the durable pattern is curl, python-requests, Go-http-client or headless Chrome strings arriving from a server subnet that has no browser on it.

Do not ignore the agent’s own telemetry. CrowdStrike found threat-actor-controlled open directories containing Claude Code session histories, ARTEX configuration files and Claude memory files during a late-September to early-October 2026 campaign against South Korean financial organizations — which means agent traces are now both an attack artifact and an investigative source. If you run internal AI agents, ship their tool-call transcripts to the same SIEM as your CloudTrail; those logs frequently explain an external anomaly before the endpoint team finds it.

Hunts you can run this week

Each of these runs against data you already have. Adjust table names, log groups and thresholds to your environment; treat the numbers as starting points rather than tuned detections.

1. Anonymous S3 principal (CloudTrail via Athena)

SELECT eventtime, eventname, sourceipaddress,
       json_extract_scalar(requestparameters, '$.bucketName') AS bucket
FROM cloudtrail_logs
WHERE useridentity.type = 'AWSAccount'
  AND json_extract_scalar(useridentity, '$.accountid') = 'anonymous'
  AND eventname IN ('GetObject','ListObjects','ListObjectsV2')
ORDER BY eventtime DESC
LIMIT 200;

2. Log mining with long-lived keys (CloudWatch Logs Insights)

fields @timestamp, userIdentity.arn, eventName, requestParameters.limit
| filter eventSource = "logs.amazonaws.com" and eventName = "FilterLogEvents"
| filter requestParameters.limit = 1000
| stats count() as calls, count_distinct(requestParameters.logGroupName) as groups by userIdentity.arn
| filter calls > 20

Pivot on any principal that surfaces, then search its request parameters for credential-shaped terms: ?password, ?AKIA, ?mongo, ?10., ?172.

3. Placeholder probe followed by a fuzzing burst

Before a burst of guesses, ARTEX probes a key that should not exist — for example zzz_nonexistent_abc.txt or Notexist-xyz — to learn what a not-found response looks like. A placeholder-name request immediately before a burst of guesses is a low-noise hunt signal.

grep -iE 'zzz_nonexistent|notexist-xyz' access.log | awk '{print $1}' | sort | uniq -c | sort -rn

Take the top talkers from that output and compare their 404 and 403 rate against the site median for the same hour.

4. Cognito sign-up burst (Logs Insights)

fields @timestamp, eventName, sourceIPAddress, userAgent
| filter eventSource = "cognito-idp.amazonaws.com"
| filter eventName in ("SignUp","InitiateAuth","RespondToAuthChallenge","GetUser")
| stats count() as attempts by sourceIPAddress, bin(5m)
| filter attempts > 25

5. Interpreter and downloader fan-out on endpoints

Get-WinEvent -FilterHashtable @{LogName='Microsoft-Windows-Sysmon/Operational'; Id=1} |
  Where-Object { $_.Message -match 'curl|wget|ncat|python|nmap' } |
  Group-Object { ($_.Message -split "`n" | Select-String 'ParentImage').Line } |
  Sort-Object Count -Descending | Select-Object -First 20

Run all five weekly, and re-run them on demand whenever an alert fires. Store each result set next to the query so you can show the difference between “never seen” and “normal for this identity.”

Response when you find one

Containment order matters more than tooling when the intruder is automated, because an agent will keep working through the next step of its plan while you open a ticket. Work the following sequence.

  1. Preserve before you revoke. Export the CloudTrail slice, proxy and access logs, and — if an internal agent is in scope — its session transcript, tool-call log and memory files. Revoking first often destroys the only record of what was attempted.
  2. Kill the identity, not just the key. Disable the IAM user, attach an explicit deny to the role or user, and invalidate existing sessions. An access key can be rotated in seconds by an automated operator; an identity that can no longer call any API cannot continue.
  3. Scope from telemetry, not memory. Pull every event for that principal from the first suspicious timestamp, across every region, including regions you believe you do not use. Enumerate the resources it touched and the data it read.
  4. Hunt the persistence an agent leaves behind. New access keys and trust policies, new Cognito app clients, Lambda functions, repository webhooks, log subscription filters pointed at external destinations. This is exactly where AWS Config earns its place: alert on configuration drift against a known baseline instead of discovering changes months later.
  5. Rebuild and verify. Recreate affected instances from clean images, rotate every secret the identity could reach, and confirm with a second, independent pass that the same access path is closed. Follow better practices for detecting and responding to security incidents for evidence handling and communications, and keep a current set of incident response and investigation tools installed and tested before you need them.
  6. Convert the incident into detections. Every noisy event you reviewed becomes a tuned rule or a documented exclusion. Write the placeholder-probe pattern and the anonymous-S3 rule into the runbook as named steps with named owners.

Finally, publish a short internal write-up with the queries that fired, the ones that should have fired, and the thresholds you changed. The teams that handle agentic intrusions well are not the ones with the most tools; they are the ones whose hunting cadence and runbook already assumed the adversary could run the next three steps before the pager went off.

Nathan Cole

Vulnerability management research, Dominion Cyber

Nathan Cole writes about vulnerability management and emerging threat research — why CVSS score alone is a poor prioritisation input, how patch operations actually get run, the mechanics of adversary-in-the-middle phishing, and the security model of AI assistants and browser extensions.

Get the weekly security brief

One email a week: what is worth patching, what is worth watching, and what is worth reading. No spam, unsubscribe any time.