What your logs could not rule out: notes on Barracuda's 44-hour intrusion and GTIG's exploit_agent
Two incident reports from mid-November are worth reading closely if you run cloud infrastructure, CI, or AI agents. We checked both against their primary sources for VSO's public incident registry. Here is what they say,
Two incident reports from mid-November are worth reading closely if you run cloud infrastructure, CI, or AI agents. We checked both against their primary sources for VSO's public incident registry. Here is what they say, what they do not, and what they suggest about logging.
exploit_agent (Google Threat Intelligence Group, 16 November)
GTIG found an exposed staging server hosting a multi-agent framework it calls exploit_agent. It integrates multiple frontier and local LLMs and runs continuous differential analysis of popular web-application software: compare the code before and after a patch announcement, identify the likely vulnerability, draft an exploit, execute it. GTIG says this is the first time it has seen vulnerability discovery and exploitation fully automated in the wild with no human involved.
The server also held exfiltrated victim data and post-compromise scripts for reconnaissance, brute-forcing, installing webshells and removing other attackers' webshells. Victim telemetry pointed to an interest in French media, entertainment and education websites. GTIG names no victims and no models, gives no counts and does not attribute the operator.
Two things a brief we received got wrong: it called the server a relay (GTIG says staging), and it called the framework fully autonomous end to end. GTIG's no-human claim covers discovery to exploitation, not the whole intrusion.
The practical point: your patch window is now measured against an automated pipeline that reads your vendor's diff. That is a patching problem, not a logging one.
Barracuda (10β13 November)
Barracuda's CISO published a detailed first-person account. The short version:
- A cloud access key issued to a vendor for an old project was never retired.
- It sat in a config file in the vendor's GitHub repo; the vendor's dev environment was compromised.
- From 04:33 UTC on 1 November to 23:31 UTC on 2 November, the attacker touched about ten cloud accounts across five platforms: cloud accounts, a data-analysis system, source-code access tokens, Kubernetes.
- It read stored secrets and decrypted IaC state files.
- It pushed short-lived branches so a build workflow would print secrets into CI job logs, encoded so that log masking would not hide them.
- It ran a read-only container mounting a host's file system to collect that host's credentials.
- It tried to create a new cloud admin identity and was blocked.
- An engineer noticed odd activity on 4 November, after it had stopped.
Barracuda says the tooling matches ARTEX, the open-source agentic pentest framework, and is explicit that this shows how the attacker worked, not who it was.
Then the sentence that matters most for this post. The attacker reached the system holding customers' XDR integration credentials. Barracuda found no direct evidence they were viewed or accessed, and says logging in that environment was not detailed enough to rule it out. It asked affected integration customers to rotate, and it is adding data-access logging where it was missing.
What this means for logging
Barracuda could reconstruct 44 hours across five platforms. Most of the intrusion was recorded. The one question it cannot answer is the one its customers care about most, and the reason is a measurement decision taken long before the attack: what the secret store logged, and at what level of detail.
No later mechanism fixes that. We work on VAP, a framework for making AI decision and action records verifiable by third parties, and its own glossary is blunt about this: an event that was never measured or recorded is permanently outside the reach of any provenance mechanism, by construction. Signing, hashing and external anchoring make a record tamper-evident. Completeness checks, such as VAP's INT-008, make it detectable when something is dropped from an anchored batch afterwards. None of them can produce a data-access event that was never logged.
So the useful questions are upstream:
- Is read access to your secret stores logged per object, or only per session? If an attacker reaches the store, "no evidence of access" and "cannot rule out access" are different answers. You want to be able to give the first one.
- Do your credentials have an end? Barracuda's key outlived the project. Short-lived, narrowly scoped credentials are Barracuda's own main lesson. Recording when a credential was issued, to whom, for what, and when it was revoked lets you answer "should this key still have worked?" later.
- Do you trust CI log masking? It matches known values. An encoded secret is a different string. Treat build logs as sensitive, restrict who can trigger workflows from new branches, and keep secrets out of jobs that do not need them.
- Where is the authoritative record of an agent's success? An agent's own report that an exploit or task succeeded is a claim. The target system's logs are the observation. If you run agents, keep both and keep them apart.
What a provenance framework would not have done
To be explicit about our own scope: VAP would not have retired that key, refused it, detected either intrusion while it ran, or shortened the 44 hours. It is post-hoc only. It would not have shown whether the XDR credentials were viewed. It produces no records of an attacker's own framework, such as exploit_agent, and it identifies no one. The current v1.2 text also does not yet specify how a credential's issue, expiry and revocation should be recorded and bound to the actions taken with it; that work is an unadopted proposal in our repository.
In the specification's words, conformance "does not warrant the correctness, fairness, or safety of the underlying AI decisions β only the integrity, completeness (at anchor granularity), and attributability of their records" (VAP v1.2 Β§1.6(c)).
Disclosures
VSO develops VAP. As of the latest published count (25 October 2026, reviewed 11 November), VAP has zero external implementations and zero Evidence Packs accepted in any proceeding; a previously published contract figure was withdrawn on 24 October and no contract, customer or revenue figure is asserted. VAP v1.2.0 is a Draft Specification and has no standing within the IETF, ISO/IEC JTC 1/SC 42, ITU-T or any other established standards body. This post was researched and drafted with Claude (Anthropic); neither incident involves Anthropic.
Sources
- GTIG, "GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold" (16 Nov 2026): https://cloud.google.com/blog/topics/threat-intelligence/agentic-threat-activity-crosses-a-significant-threshold
- Barracuda, "Forty-four hours: Anatomy of a machine-speed intrusion" (11 Nov 2026): https://blog.barracuda.com/2026/11/11/anatomy-of-a-machine-speed-intrusion
- Barracuda Trust Center, "Barracuda XDR Integration Credentials" (10 Nov, updated 13 Nov 2026): https://trust.barracuda.com/security/barracuda-xdr-integration-credentials
- VSO incident registry: https://veritaschain.org/incidents/
- VAP v1.2 specification (Β§1.6, Β§4.1.7, INT-008, Appendix A): https://github.com/veritaschain/vap-spec/blob/main/spec/v1.2/VAP_Framework_Specification.md
Longer version on Medium: https://medium.com/@veritaschain/the-access-nobody-logged-2fd7a1c44eee
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes β full credit and traffic to the original publisher.