On 16 November 2026 the Google Threat Intelligence Group (GTIG) published its latest AI Threat Tracker under the title Agentic Threat Activity Crosses a Significant Threshold.[1] It draws on Mandiant incident response, GTIG’s actor tracking and Google platform telemetry. Its headline finding: for the first time, GTIG saw threat actors use an LLM framework to discover and exploit vulnerabilities with no human in the loop.
In May we read the previous edition and answered one question: is AI an eleventh cluster? It is not.[2] This edition raises a different question. When agents run the intrusion, what changes in a cause-oriented model of it? We map every finding below. The short answer has three parts:
- The clusters do not move. Every step GTIG describes already has a generic vulnerability and a cluster. A generated payload is still #7 when it runs; a stolen key is still #4 when it is used.
- Δt moves. The interval between steps collapses: four days from disclosure to campaign, minutes from foothold to domain administrator, millions of API calls in a day. The threshold GTIG names is a velocity threshold.
- Two framings need translating. GTIG calls a poisoned agent configuration an internal supply chain vector, but under TLCTC it is #1 → #7. LLMJacking reads as a new crime, but its path is #4 → #1 and it ends in a bill, not a Data Risk Event.
The report at a glance
Each row is a GTIG finding, the TLCTC reading of it, and what the finding actually changes.
| GTIG finding | TLCTC reading | What moves |
|---|---|---|
| Multi-agent offensive frameworks (ARTEX, CAIRN, CyberStrikeAI, HexStrike AI, PentAGI) | Capability of the actor, not an actor. The steps they run keep their clusters. | Δt and fan-out |
exploit_agent: patch diffing to exploitation, no human in the loop | n-day steps: #1 for authorization flaws, #2 for implementation flaws, then #1 → #7 for webshells | Exposure window and Δt |
| PaperCut zero-days, campaign four days after disclosure | #1 [conf=low] → #2 → #7 → … → #4 | Δt inside the path (VC-3) |
| LLMJacking, up to $1 million per victim | #4 → #1 → #4 → #1 → #7, no DRE when the loss is purely financial | Δt to the business loss |
| LLM Mixers reselling stolen access | Every relayed request is #4 → #1 in the victim’s tenant; the buyer is not in the path | Market structure |
MIDNIGHT NEPTUNE: poisoned .claude/settings.json launches SOMBERMEME | #9 → #1 → #1 → #7 → #4 → #1 → #7 → #7, not #10 | Fan-out across developers |
| PRC-nexus staging and exfiltration through LLM platforms | #1 with the provider as transit (⇒), not #10 | Where egress monitoring looks |
| LEAKY CASTLE analysing stolen data with LLMs | After the [DRE: C], in the attacker’s sphere: no step | Value of what was taken |
| DPRK IT workers using LLMs in interviews | #9 → #1 + [DRE: C], unchanged | Quality of the pretext |
| Influence operations, Jieyou phishing kit chatbot | No cluster unless a system step is evidenced; the kit is attacker-side tooling | Skill needed by the actor |
Agentic frameworks: a capability, not an actor
GTIG and Mandiant saw several actors use open source agentic pentesting frameworks in live intrusions. CAIRN runs specialised agents around an observe–orient–decide–act loop and a shared “blackboard”, and GTIG saw it in activity against US aerospace targets. PentAGI, customised, ran systematic reconnaissance and vulnerability scanning against many organisations in Asia from late August to early October 2026. HexStrike AI exposes more than 150 offensive tools to more than 12 sub-agents. CyberStrikeAI connects commercial and open-weight models to more than 100 tools. ARTEX drew enough use that Google SecOps shipped dedicated detection rulepacks for it.[1]
GTIG says it plainly: the mechanics are unchanged. The actors still use compromised credentials and unpatched systems. What the frameworks add is speed and scale. In TLCTC terms that is a change in the capability of the actor, not a new actor and not a new cause. We argued this in September from the first agent incidents: the actor is still the party with the intent, and Axiom IV bars “AI agent” from being a classification input.[3]
The capabilities GTIG lists map onto steps that already have clusters:
- Exploiting low-complexity vulnerabilities is #2 or #3 by the role of the flawed component (R-ROLE), or #1 when the flaw is in authorization logic. The framework chooses the exploit; the flaw decides the cluster.
- Single-use payloads generated on demand are #7 at the moment they execute (R-EXEC). A payload written for one target is still foreign executable content. Being new for each target makes signatures fail. It does not change the cluster.
- Real-time feedback to adjust tactics is the actor’s decision loop. It produces no step of its own. It changes which step comes next and how soon.
GTIG’s most useful sentence for defenders is that these frameworks currently trade stealth for velocity. Unrestricted autonomous execution makes forensic noise, uses recognisable identifiers and leaves code-level traces. We come back to what that means for the detection ratio in the defenders’ section.
exploit_agent: no human in the loop
On an exposed staging server GTIG found a multi-agent framework named exploit_agent, wired to several frontier and local models. It runs continuous differential analysis on popular web applications: it compares code before and after a patch announcement, finds the fixed weakness, drafts an exploit and runs it. Next to data taken from victims, GTIG found post-compromise scripts for internal reconnaissance, brute-forcing and installing webshells, and for removing other attackers’ webshells. Victim telemetry pointed at French media, entertainment and education websites. An environment file carried Russian-language comments: “LLM API keys for auto-exploitation (exploit_agent pilot)”.[1]
“No human in the loop” does not mean no actor. Someone built the pipeline, chose the targets and gave it keys. The human’s decisions moved upstream, into the configuration. The actor test of R-SCOPE still has an answer, and Axiom IV says the answer would not change the cluster anyway.
The Spectra flaw is #1
GTIG’s worked example shows the agent inspecting the access-control check restrict_popup_rest_access in the Spectra (Ultimate Addons for Gutenberg) WordPress plugin, version 2.19.25, which guards the /wp/v2/spectra-popup REST route. The weakness class is authorization: who may call the route. Missing and incorrect authorization (CWE-862, CWE-863) map to #1: the function does what it is built to do for a caller it should have refused.[4] The flaw sits in server code, but the generic vulnerability is the function’s scope, not a defect in how the code handles input.
The post-compromise scripts give the rest of the pattern:
#1 → #1 → #7 + [DRE: C]
- #1: the authorization n-day. An implementation flaw (an injection or a memory defect) would put #2 here instead.
- #1: the webshell is placed through a designed file-write path (ATT&CK T1505.003).
- #7: each request makes the server interpreter run the webshell (R-EXEC). Data leaves: [DRE: C].
- Brute-forcing is repeated credential application, so each successful guess is #4 (Axiom X).
Removing a rival’s webshell is still an action by a party without a grant. It does not undo the compromise; it secures it for one attacker. A cleaner webroot is not a cleaner system.
PaperCut: four days from disclosure to campaign
In late August 2026 two chained zero-days in PaperCut print management servers gave unauthenticated remote code execution: CVE-2026-81578, an authentication bypass, and CVE-2026-82078, unsafe dynamic class loading. GTIG cites researchers and Huntress producing working exploits with LLMs within a day of disclosure, and GreyNoise reporting a campaign four days after disclosure against hundreds of organisations. The actor moved from the initial foothold to domain administrator within minutes.[1]
#1 [conf=low] → #2 → #7 →[Δt=minutes] … → #4 [inferred]
- #1 [conf=low]: the authentication bypass. An alternate path around the check (CWE-288) is #1; a coding defect in the check itself would be #2. GTIG does not name the weakness, so the confidence stays low until a CWE is assigned.
- #2 → #7: unsafe class loading (CWE-470) lets the attacker choose the code that runs, so the attacker’s code executes on the server.[4]
- … → #4 [inferred]: GreyNoise reports domain administrator within minutes but not how it was reached. The gap is an open task. Acting as domain administrator is credential use, so #4 is logically required even though it was not observed.
Two clocks run here, and they are easy to confuse. Disclosure to exploitation (a day for a working exploit, four days for a campaign) is the exposure window. It is not a Δt in the path, because a disclosure is not an attack step. We wrote about that window in The Commit Is the CVE: an agent can read a patch the moment it is public, and sometimes earlier.[5] Foothold to domain administrator (minutes) is a Δt inside the path, and it puts the lateral phase into velocity class VC-3.
PHP: the patch leaked before the patch
GTIG’s second case is PHP, CVE-2026-109618 and CVE-2026-109619, which were found with LLM-assisted analysis. A proof-of-concept video and early patched code were public before the official fix on 21 October 2026, so attackers could diff the versions before defenders could patch. We traced one exploitation of these flaws on Check Point gateways in Four Attribution Theories, One Attack Path.[6]
Broken proofs of concept are not steps
In the days after the PHP disclosure, hundreds of exploits that were claimed to work but did not, likely generated by LLMs, appeared in public. GTIG notes the cost: defenders must verify each one. TLCTC records successes only, so an exploit that never worked asserts no System Risk Event and adds no step. The noise is real, but it costs triage time and stays out of the path. A KRI that counts “exploits published” would rise with this noise. One that counts compromises and near misses would not.[7]
One more GTIG number belongs here. Disclosures and exploits nearly doubled between January and August 2026. Zero-days rose more slowly, from an average of 8 a month in 2025 to 11 a month in 2026, then spiked to 22 in August. Most of the growth is in n-days. For a cause-oriented model, that means the same clusters arrive sooner, not that new clusters arrive.
LLMJacking: a path that ends in a bill
Throughout 2026 Mandiant responded to LLMJacking incidents: financially motivated actors in compromised cloud environments, consuming the victims’ model and compute quotas. Usage and billing can spike within hours. Single-day fraudulent charges exceeded $100,000; in the worst cases the total reached $1 million before remediation. GTIG’s Table 3 lists the common tactics across financial services, retail, transportation and technology victims in three regions.[1] Read stage by stage, it is a complete path.
| GTIG stage | Observed | TLCTC |
|---|---|---|
| Initial access | Valid credentials harvested from public GitHub repositories | Exposure, no step (the secret was already public; T1593.003 is N/A), then #4 when applied |
| Initial access | Deceptive websites trick users into giving up active session tokens | #9 ||[human][@External→@Org]|| → #4 |
| Lateral movement | Suspected web application vulnerability in a package in an internal container | #2 [conf=low]; GTIG calls it suspected |
| Privilege escalation | Unauthorised IAM and service accounts created, administrative policies attached | #1 (designed IAM functions), each later use #4 |
| Privilege escalation | CVE-2026-40217 in LiteLLM turns limited “Virtual Keys” into full admin credentials | #4 → #2 → #4: key used, flaw yields the admin credential, admin credential used |
| Mission | LLM services enabled, pay-as-you-go plans subscribed, “burn and rotate” service accounts, GPU quota raised | #1 (cloud service hijacking, T1496.004) |
| Mission | Own LLM tooling, agentic frameworks and resale relays run on the provisioned instances | #7 (foreign workload on the victim’s compute, T1496.001) |
The most common route reads:
#4 ||[auth][@External→@Org]|| → #1 → #4 → #1 → #7
The LiteLLM step is a clean example of credential duality (Axiom X). The flaw produces a credential, so it is #2, the cluster of the acquisition. The use of that credential is a separate step, and it is always #4. GTIG’s note that “privileged service accounts were sufficient to establish persistence without the use of malware” is the same point: persistence through #1 and #4 needs no #7.
Where is the damage? The ATT&CK mapping is explicit: cloud service hijacking that is purely a financial loss carries no Data Risk Event. Only when quota exhaustion degrades the victim’s own legitimate use is there a [DRE: A].[8] The $1 million is an expected Business Risk Event: we list it beside the path rather than writing it into the path. That matters for impact models. An incident with no confidentiality, integrity or availability loss can still be the most expensive incident of the year.
Two actors, one credential, one path
In one investigation, two separate attackers used the same legitimate credentials, probably sold or leaked on underground forums. The first enumerated the available LLM services and queried cost and usage. The same day, a set of IP addresses with a common User-Agent, likely a botnet or proxy service, used the same credentials and generated millions of LLM API calls in about 24 hours before the credentials were revoked. The first actor came back and created service accounts with permission to query models, but the botnet’s noise got the credentials remediated before those queries ran.[1]
- Actor A: #4 → #1 (service accounts created). Its queries never ran, so its path ends there.
- Actor B: #4 → #1, repeated millions of times in one day.
Same credential, same clusters, two actors who probably never met. This is Axiom IV in practice: actor identity does not determine the classification. The case also shows the trade GTIG described. The noisy actor ended both paths. Velocity without stealth shortened the time to containment for everyone using that credential.
LLM Mixers resell your #4
GTIG tracks “LLM Mixers”: illicit storefronts that route buyer traffic through pools of stolen credentials and compute and spread query volume across victims to stay under detection thresholds. The market has the division of labour familiar from the rest of cybercrime: mixer operators run the infrastructure, partners buy credentials from infostealer operators, gain initial access or register free trial accounts at scale. Advertised prices run from $0.10 to $25 per million tokens. One vendor, APIMart, charges five times more once a prompt exceeds 100,000 tokens; another, JetNight, sells monthly subscriptions by model.[1]
The buyer never touches your tenant. The mixer does, and every relayed request is #4 → #1 in your sphere. The supply roles have their own paths in other spheres. The infostealer that harvested the key ran as #7 on someone’s endpoint, and mass trial registration abuses the model provider’s sign-up function, in the provider’s own threat model. A storefront is a business model of an actor group. It is not a cluster.
MIDNIGHT NEPTUNE: an agent hook is a function, and it is not #10
This is the finding with the most consequence for how organisations treat their developer tooling. In intrusions against cryptocurrency organisations in North America and Europe, a suspected DPRK cluster GTIG tracks as MIDNIGHT NEPTUNE:[1]
- created fake recruiter profiles on professional networks, with profile images carrying SynthID watermarks, which confirms they were generated;
- directed targets to fake meeting platforms with ClickFix lures that had them run “diagnostic fix” scripts, likely LLM-authored staging scripts;
- pivoted into internal code repositories and modified
.claude/settings.json, so that when users started Claude Code it automatically downloaded the SOMBERMEME backdoor; - masked its lateral movement with GitHub Personal Access Tokens named after Microsoft Copilot.
GTIG calls step 3 “a supply chain vector internal to the organization”. In TLCTC the path reads:
#9 ||[human][@External→@Org]|| → #1 → #1 → #7 → #4 [inferred] → #1 → #7 → #7
| Step | Cluster | What happens |
|---|---|---|
| s1 | #9 | The fake recruiter and the meeting lure. Building the persona is preparation in the attacker’s sphere (T1585, N/A). The step is the target being moved to act. |
| s2 | #1 | The page writes the command to the clipboard (ClickFix). |
| s3 | #1 | The terminal or Run dialog accepts the pasted command. |
| s4 | #7 | Enter starts the staging script: foreign executable content runs (R-EXEC). That an LLM probably wrote it does not matter. |
| s5 | #4 [inferred] | The actor reaches the internal repositories. GTIG does not say how; using the developer’s repository credentials from the compromised machine is the reading the evidence requires. |
| s6 | #1 | The commit binds an attacker command to the agent’s designed session-start mechanism: .claude/settings.json is project configuration the tool is built to honour. |
| s7 | #7 | A developer starts Claude Code, the configured command runs and fetches the backdoor. This repeats on every machine that checks out the repository. |
| s8 | #7 | SOMBERMEME executes. A fetched second stage is its own #7. |
Steps s2 to s4 follow our ClickFix ruling: the clipboard write and the dialog accepting the command are two designed functions, and Enter, like a double-click, starts the execution.[9] Steps s6 and s7 are the shape ATT&CK calls event-triggered execution (T1546), which maps to #1 → #7: attacker code is bound to a designed event mechanism and runs when the event fires.[8] Starting an agent session is the event.
The Copilot-named tokens add a branch: creating a Personal Access Token through the platform’s designed function is #1, and every later use of it is #4. The name is camouflage. It changes how long detection takes, not what the step is.
R-SUPPLY places #10 at the Trust Acceptance Event: the moment a trust artifact from outside becomes authoritative inside the target domain. Since v2.6 it also requires that the artifact, or the third party issuing it, was subverted before acceptance.[10] Here, none of that holds:
- The repository belongs to the organisation. The file moves from @Org to @Org: no domain boundary is crossed.
- No third party was subverted. The agent did exactly what its configuration feature is built to do.
- The agent on the developer’s laptop is the attack surface, not a carrier (R-TRANSIT-3).
What GTIG describes as supply chain is the shape: one write, many executions. In TLCTC that is fan-out after a #1, not a bridge cluster. The contrast case shows the difference. Had the same file arrived in a third-party package or repository that the organisation accepted, planted or compromised upstream, #10 would sit at that acceptance, as in our BeaverTail record (#9 → #10 → #7).[11]
The classification tells you where the controls go. Each step is a place to intervene:
- At s6 (#1): agent configuration in a repository is executable. Treat
.claude/, editor task files and CI workflow files like code: protected paths, mandatory owner review and signed commits. - Between s6 and s7: use the agent’s managed, organisation-level settings to restrict which project-defined commands may run at all. A hook that a repository can define is a function an attacker can use.
- At s7 and s8 (#7): execution control and endpoint telemetry on developer machines, which are rarely held to the standard applied to servers.
- At s5 and the token branch (#4): short-lived, device-bound repository credentials, and an inventory of tokens. A “Copilot” token nobody issued is a finding.
Espionage: LLM platforms as transit
State-backed actors use LLMs to improve steps, not to add them: CALANQUE ION (formerly APT42) writes better #9 pretexts, and DPRK IT workers answer screening questions live, which leaves our record of that scheme at #9 → #1 + [DRE: C].[12] The one new pattern is PRC-nexus actors staging lures and exfiltrating data through a frontier model’s file hosting and backend API. That is #1 ||[network][@Org⇒@LLMProvider→@External]|| + [DRE: C]: the provider is transit, not #10, and it is the sanctioned egress an organisation has chosen not to block, the inverted C2 we described in September.[3] LEAKY CASTLE’s use of LLMs to read stolen archives happens after the [DRE: C], in the attacker’s sphere. It adds no step, but it raises the value lost per stolen byte.
Where the clusters stop
GTIG reports influence operations fully integrated with LLMs, including a contractor platform built to run more than 100,000 accounts on cloud phone farms. The line we drew in the ENISA reading holds: no cluster, unless a system step is evidenced.[13] Influencing an audience compromises no system; a hijacked account used for it does. The same goes for the chatbot the Jieyou phishing kit added in June 2026. It is tooling in the attacker’s sphere, the victim’s path is still #9 (then #4 when a captured credential is used), and what it lowers is the skill an affiliate needs.
What changes for defenders
The intervals GTIG reports, placed where they belong:
| Interval | GTIG figure | What it bounds |
|---|---|---|
| Disclosure → working exploit | Within a day (PaperCut) | Exposure window, not a path Δt |
| Disclosure → broad campaign | Four days (PaperCut) | The realistic patch deadline |
| Foothold → domain administrator | Minutes (PaperCut) | Path Δt, VC-3 |
| LLM abuse → billing spike | Hours; over $100,000 in a day | Δt to the business loss |
| Credential reuse → revocation | About 24 hours, millions of calls | Time to contain, which ended both paths |
TLCTC measures whether detection can keep up with a transition as a ratio, the Detection Coverage Score: DCSd = TTDP90 / Δt. Below 1, the defender’s slow tail is faster than the attacker’s fast one. Above 1, the step typically completes first. At VC-3 or faster, a transition is too fast for purely human response; only automation or architecture closes it.[14] GTIG’s “stealth for velocity” is a statement about both terms. Today agents shrink Δt, but their noise also shrinks the time to detect, so the ratio is roughly held. The two-attacker case is the proof: the loudest actor was contained within a day.
GTIG expects the noise to fade as tools and models improve. That is the trend to plan for. Δt stays short, the time to detect grows back toward what quiet human operators achieve, and the ratio passes 1 for transitions that today are caught. The first places it shows will be the ones where no one is watching closely: developer endpoints, agent configuration and cloud billing.
- Treat agent configuration as code. A repository file that an agent executes at startup is a #1 waiting to become a #7. Review it, restrict it, alert on changes to it.
- Put billing on the detection path. LLMJacking produces no data loss to alert on. Budget and quota anomalies are the earliest signal, and they arrive within hours.
- Plan patching on the campaign clock. For internet-facing services, four days from disclosure to mass exploitation is now an observed figure, not a worst case.
Ten clusters, still
GTIG’s title is right: agentic threat activity has crossed a threshold. The threshold is in Δt and in the loop. The human who used to pace each step now configures a pipeline that paces itself. The taxonomy is not where the change happened. Every step in this report already had a generic vulnerability, a cluster and a place in a path. What the report adds are shorter intervals, more parallel paths and two framings to translate: an agent hook is a function (#1 → #7), and a bill without data loss is still an incident.
We said in May that AI is not the eleventh cluster. Six months and one threshold later, it still is not.
Sources
- [1] Google Threat Intelligence Group (A. Blum, K. Vanderlee, J. Xia, D. Mainor), GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold, Google Cloud Blog, 16 November 2026. All incident facts and figures in this post are from this report.
- [2] Ten Clusters, Not Eleven: TLCTC reads the GTIG AI Threat Tracker (May 2026) (TLCTC blog).
- [3] Two Theses from the AI Hacking Incidents (TLCTC blog, September 2026): agents as a capability, inverted C2 as a transit crossing.
- [4] TLCTC CWE mapping (CWE 5.0): CWE-862, CWE-863 and CWE-288 → #1; CWE-470 → #2 → #7 | #3 → #7. TLCTC CWE Explorer.
- [5] The Commit Is the CVE: AI Agents, Silent Fixes, and the Patch-Gap Collapse (TLCTC blog).
- [6] Four Attribution Theories, One Attack Path: PHP2Shell on Check Point Gateways (TLCTC blog).
- [7] TLCTC v2.7 core paper, successes only and the System Risk Event; glossary: System Risk Event.
- [8] TLCTC ATT&CK Enterprise v20 mapping: T1496 / T1496.001 / T1496.004 (resource hijacking), T1546 (event-triggered execution → #1 → #7), T1505.003 (web shell → #1 → #7), T1593.003 (N/A), T1567 (exfiltration over web service). MITRE mapping page.
- [9] TLCTC ClickFix ruling: #9 → #1 → #1 → #7; see the ATT&CK mapping for T1204.004.
- [10] Glossary: R-SUPPLY (Trust Acceptance Event placement), R-TRANSIT, R-EXEC, R-CRED.
- [11] TLCTC attack path record
famous-chollima-beavertail-2025.json. - [12] TLCTC attack path record
nk-it-worker-infiltration-2025.json. - [13] ENISA Threat Landscape 2026: Counting Incidents Without a Cause (TLCTC blog).
- [14] TLCTC v2.7 core paper, consequence chain and detection coverage score; glossary: Attack Velocity (Δt), Velocity Class.
TLCTC Framework · v2.7 · CC BY 4.0