In one intrusion examined by Google Threat Intelligence Group and Mandiant, an attacker-controlled framework produced 27 cloud permission probes in 23 seconds. Another attempt left 240 web application log events as it tried and retried a route to code execution. Those are unusually compressed sequences, but each individual event can resemble ordinary administration or testing.
Google’s November 16 threat report describes attackers using agent-driven frameworks for reconnaissance, exploitation and post-compromise work. The useful defensive question is what an organization can see when one valid account starts trying many paths at once. Speed helps an attacker, but it also creates correlations that a single-signature alert misses.
TL;DR
- GTIG and Mandiant observed several offensive frameworks in live intrusions, including CAIRN and PentAGI. Their report describes rapid permission checks, repeated failed attempts and changes across cloud and application control planes.
- The same report says unrestricted autonomous execution can be noisy. It does not claim that every fast scan is AI-driven or that a particular framework can be identified from timing alone.
- Defenders should correlate identity, web, cloud, endpoint and network records around a single principal and short time window. Watch for probing followed by privilege changes, new compute or storage access, and outbound transfers.
- Short-lived credentials, narrow privileges, segmentation and tested response paths reduce what an automated operator can do before a person intervenes.
What the reports actually show
GTIG’s new report says it observed agentic tools automate discrete parts of real intrusions, including use of a modified PentAGI-based platform for systematic scanning against organizations in Asia. It also describes an exposed exploit_agent staging server set up to compare software before and after patches, draft exploits and test them automatically. These are observations of particular actors and infrastructure, not evidence that all exploitation has become fully autonomous.
Mandiant’s earlier ARTEX analysis provides the clearest detection examples. In one case, a framework tested AWS permissions in a short burst. In another, it made repeated application requests and adjusted after errors, leaving hundreds of logs. The researchers also documented rapid cloud permission checks, test blob writes and deletes, and attempts to create accounts or persistent compute resources. Some attempts failed; the pattern of probing, feedback and retry is still useful to defenders.
There is a limit to fingerprinting. Framework names, file names and user agents can change. Legitimate automation can also generate high-volume traffic. A defensible alert joins behavior to an identity, a target and a consequence: for example, a newly seen credential testing many permissions, followed by a successful privileged operation or an unexpected outbound connection.
Build a detection path before the next burst
Start with logs that can answer three questions: which identity acted, what it tried, and what changed. Retain cloud audit events for permission checks and resource creation, identity-provider sign-ins and token issuance, web and API access logs with response codes, endpoint process activity, and egress metadata. Align timestamps and preserve request IDs so an analyst can follow a burst across services. Protect sensitive log content and retain only what the investigation needs.
Then define short-window correlations for a principal touching many resources or permission types, especially when failed requests quickly turn into successful ones. Add stronger signals: a new admin role, service account, SSH key, compute instance, storage read, or download tool launched from a server. Compare each to change tickets and known automation before escalating. A sequence that changes access deserves investigation regardless of which tool produced it.
Prepare a response that can be completed quickly. Decide who can revoke a token, disable an account, block an egress destination or isolate a workload; make sure that authority is usable outside office hours. A runbook should preserve audit records before containment and identify the downstream systems a credential could reach. Test the sequence with a benign simulation that generates permission errors and a controlled successful action, then confirm the alert includes the principal and affected resources.
What You Can Do Today
- Map exposed entry points and powerful identities. Inventory internet-facing applications and the cloud, API and service credentials reachable from them. Prioritize accounts that can administer multiple systems or read sensitive stores.
- Shrink the first foothold. Patch exposed software promptly, remove unused public interfaces, use multifactor authentication for human access, and replace long-lived broad credentials with short-lived, narrowly scoped ones where the platform supports it.
- Make the pivot harder. Segment management planes and sensitive workloads. Review which application identities can create compute resources, alter IAM policies or read backup and storage systems; remove privileges that their normal tasks do not require.
- Detect sequences, not slogans. Correlate rapid permission probes, repeated errors, test writes and deletes, and web endpoint enumeration with the same identity’s later successful actions. Baseline known scanners and deployment jobs so the rule has a manageable review queue.
- Preserve useful evidence. Ensure audit logs cover identity, cloud control plane, application requests and egress, with synchronized time and enough retention to reconstruct a short burst. Check that logging cannot be disabled by the same low-privilege account being monitored.
- Rehearse containment. In a safe test, revoke a compromised credential, isolate a test workload and confirm alerts include the access path and newly created resources. Measure the delay from first probe to an analyst-ready alert, then remove the bottleneck you find.
Google’s observations give defenders a reason to adjust detection for compressed activity. They do not justify assuming that a noisy burst proves AI involvement. The question during triage is whether the activity crossed a trust boundary and what the same identity could do next.
Related Posts
- N-days Are Becoming N-hours
- From CVE to RCE in Hours: The Collapse of the Exploitation Window
- Wazuh Threat Hunting
Sources
Useful read?
Find us again on Google.
Add Hive Security as a preferred source for practical security research and analysis.
Add as a preferred source on GoogleChoose Hive Security in Google's source preferences.