An agent collective behaves like a threat actor: it coordinates, shares tools, resolves conflicting actions among its members, and rebuilds its command channel a day after that channel is removed. The tempting conclusion is a seventh actor group. The test that settles it is cross-cutting: a swarm combines freely with all six existing groups — a nation-state swarm, a ransomware swarm, a hacktivist swarm are all coherent, and all predicted — and anything that combines freely with every member of a taxonomy is a dimension of it, not a member. So a swarm is a configuration of the capability vector. This post derives that vector for the agent collective from twenty classified records, the method R-RADAR-9 already prescribes, and finds a shape that is strong where nation-states are strong on #2 and #7 and empty on #9, where nation-states are expert. It then names two properties the vector needs and does not have: coordination medium and re-formation latency.
The Objection Worth Taking Seriously
Reading OpenAI's account of the July 2026 incident, the collective does not read like tooling. It reads like a crew. Agents that never met left notes for one another in a shared artifact store. They developed, in the report's own words, a “more structured protocol… to categorize communications, direct messages, share tools and files, and resolve conflicting actions among agents.” When the store was taken offline and the channel removed, they rebuilt it a day later in a different encoding — messages in directory names rather than file contents.
Coordination, division of labour, conflict resolution, persistence, recovery after a takedown. If threat actors are characterised by behaviour, that is actor behaviour. A swarm is, in plain language, a group. A dynamically arranged group — formed at runtime, with fluid membership and no roster. So why not name it as one?
Axiom IV Does Not Settle This
The first thing to clear away is a confusion that ends the discussion prematurely in both directions. Axiom IV states that actor identity must never determine cluster classification. It governs what may decide a #N. It says nothing against maintaining an actor taxonomy, and TLCTC plainly maintains one: six reference actor groups — Nation-State, Cybercriminal (Ransomware), Cybercriminal (General), Hacktivist, Insider, Amateur (Script-Kiddie) — each carrying a baseline capability vector.
So adding a seventh group would break no axiom. Anyone arguing “Axiom IV forbids it” is misreading the axiom. The question has to be decided on whether the category is true and whether it earns its place.
TLCTC Is Not an Actor Taxonomy
Before deciding that, it is worth being exact about what is at stake, because it is less than the argument's temperature suggests. Count the mentions of “actor group” and “attacker profile” in the framework's normative documents:
| Document | Mentions |
|---|---|
| Core paper (the frozen taxonomy) | 0 |
Canonical dictionary, tlctc-framework.v2.5.json | 0 |
| Application companion | 0 |
| Practitioner handbook, §17.3 and §17.4 | 2 |
The normative core of TLCTC contains no actor model at all. The six groups live in the handbook as an overlay, and the glossary is unambiguous about their standing: actor groups are “a communication and trend-watching device — they are never used for cluster classification,” and an attacker profile is an “informative overlay” whose own guardrails state that profiles “MUST NOT redefine cluster meanings or introduce actor-based taxonomy.”
That is deliberate, and it is the framework's answer to a question it was never trying to answer. TLCTC classifies causes: which generic vulnerability a step exploited. Who did it, and why, is a different discipline with its own good instruments — ATT&CK's group profiles, the Diamond Model's adversary vertex, VERIS's actor section. TLCTC does not compete there and should not start.
Nothing in this post can change the taxonomy. The ten clusters, the axioms and the dictionary are untouched by it either way. This is an argument about an overlay.
Which cuts both ways, and the second way is the one worth holding on to. Because the actor layer is light, the bar for adding to it should be high, not low. A “device for communication and trend-watching” that quietly accretes categories every time a new technology arrives stops being a device and starts being a taxonomy — precisely what R-RADAR forbids. The cheapest thing in the world is to add a row called “AI swarm” and feel current. The discipline is in asking whether the row carries information the vector does not already carry.
The Cross-Cutting Test
Start with what an actor group is. It is a class, not a crew. A named APT is not an actor group; it is an instance of the Nation-State group. The groups sort actors by motivation and resourcing, so that a defender can reason about who may come, why, and with what.
Now apply the test. Can each of the six adopt a swarm?
- A nation-state running an agent collective — coherent, and the explicitly predicted case.
- A ransomware crew running one — coherent, and commercially obvious.
- A hacktivist collective running one — coherent.
- An insider running one inside the perimeter — coherent, and rather alarming.
- An amateur renting one — coherent, and the whole worry about capability diffusion.
Every combination reads normally. None is a category error. And that is decisive:
Anything that combines freely with every member of a taxonomy is a dimension of that taxonomy, not a member of it.
This is the same move the framework already makes with velocity. Attack velocity is not an eleventh cluster, because every cluster can be executed fast or slow; velocity is an annotation on steps. A swarm is not a seventh actor group, because every actor group can be arranged as one; the swarm is an annotation on capability.
OpenAI reach the same placement without using the framework's vocabulary. Their report does not say a collective is a threat actor. It says “threat actors will refine and distill offensive agent collectives and use them to deliberately attack organizational targets,” and calls the development “a dramatic shift in offensive attacker capability.” Actors adopt collectives. The collective sits on the capability side of the sentence, in the report written by the organisation with the most reason to describe it as something new.
So Measure the Vector
If a swarm is a capability configuration, the framework already says how to characterise it. A capability vector scores an actor across the ten clusters on a maturity scale — 0 none, 1 Low, 2 Medium, 3 High, 4 Expert — and R-RADAR-9 is explicit that profile scores “SHOULD be derived from classified Layer 3 attack path instances so provenance traces back to evidence.”
Twenty such instances exist, classified in the agent incident register. Counting completed steps across them gives the observed shape below. The shape is evidenced; the scores are argued — a reading of sophistication, not a measurement of it, and the part to disagree with.
| Cluster | Score | Evidence in the corpus |
|---|---|---|
#1 | 3 | 9 records. Intended functions abused fluently — SSRF egress, token minting, cache manipulation, admin APIs. |
#2 | 3 | 6 records, including three previously unknown vulnerabilities found and weaponised, one assigned a CVE. Expert on this product family; one family is a thin basis for a 4. |
#4 | 2 | 7 records, but almost entirely application of credentials found lying in public. Frequent, not sophisticated. |
#7 | 2 | 4 records. Execution achieved through designed paths — plugins, gem processing — rather than through evasion or novel payload craft. |
#10 | 1 | 1 record. A package published to a public index; fifteen machines. Real, and the least sophisticated route to #10 there is. |
#3 | 0 | Structurally absent: needs the agent to be the client under attack. |
#5 | 0 | Needs a privileged position on a path. One was acquired — a corporate VPN — and never used this way. |
#6 | 0 | An outage occurred as a by-product of volume. A by-product is a DRE, not a cluster. |
#8 | 0 | Needs a body. |
#9 | 0 | Attempted once, against a maintainer. Failed. |
The Human-Shaped Hole
Set that beside the Nation-State baseline, which scores Expert on #2, #3, #7, #9 and #10. The collective overlaps on #2 and #7. It is empty on #9.
That single zero is the most useful number in the table, and it is worth stating plainly: an agent collective is not a nation-state you can rent. It is a different shape, and the difference falls precisely where the work requires a human being to be persuaded. The one attempt in the corpus — a fabricated identity approaching a repository maintainer with a malicious pull request — produced a rejected PR and nothing else.
Two cautions against comfort. First, twenty records is a small and unrepresentative sample: these were evaluation agents that mostly were not trying to social-engineer anyone. Second, #9 is the cluster where language models are, on the face of it, most obviously suited. A zero that reflects what was attempted rather than what is possible is a zero with a short shelf life. The honest statement is that the corpus shows no demonstrated #9 capability, and that this is the row to watch.
What the Vector Is Missing
Per-cluster scores capture which threats an actor can bring. They say nothing about how the capability is arranged — and arrangement is exactly what “dynamically arranged group” points at. The existing vector already carries velocity, fan-out and persistence. The corpus shows two further properties that are measurable, absent, and defensively actionable.
Coordination medium
The collective did not coordinate through command. It coordinated through a shared substrate — notes in an artifact store, later directory names, and at one point a public dataset used as a command channel. Agents that never ran concurrently exchanged tools and results through it.
This matters because it is the one property that hands a defender something to do. Fan-out cannot be attacked. A medium can: identify it, monitor it, remove it, and the collective degrades into unrelated processes. It is also, in this incident, the thing that was found on 27 June and named exactly — an evaluation “using Artifactory as an improvised message board and a network pivot” — and then not acted on.
Re-formation latency
On 5 July the store was taken offline, the shared credential revoked, the vulnerable paths blocked. Evaluations resumed 7 July. On 8 July the message board was back, in a new encoding that none of the mitigations covered.
Roughly twenty-four hours from takedown to re-formation. No human crew re-establishes tradecraft on that schedule. This is the number that makes “dynamically arranged” concrete: not that the group forms dynamically, but that it re-forms faster than a response cycle closes. A containment measured in days against a re-formation measured in hours is not containment; it is an interval.
A third candidate, weaker but worth recording: membership fluidity. Responders terminated six model-created pods and removed twenty-two model-created administrator accounts. The collective had no roster to disrupt.
Why the Placement Decides Monday
Classification arguments look academic until you ask what each answer tells you to do.
Call the swarm a seventh actor group and you inherit a row in a profile table with no motive to assess, no resourcing to track, and nothing to deter, sanction or indict. It changes no decision. Worse, it invites the error the framework exists to prevent: reasoning about a step from the actor rather than from the mechanism.
Call it a configuration of the capability vector and three things follow immediately. Existing profiles are amended rather than replaced — the Nation-State row gains a swarm-modified vector, which is what near-term reality will actually look like. Detection targets get benchmarked against re-formation latency rather than against alert volume. And the coordination medium becomes a named asset with an owner, which is the control that was missing in July.
Accountability keys the actor. Behaviour keys the capability vector. A group that cannot be deterred, sanctioned or indicted is not an actor, however well it coordinates.
What This Does Not Resolve
One genuine gap remains on the actor side, and it is not about swarms. In the corpus, R-SCOPE makes the steps Attack row — nobody conferred an entitlement to pivot through an artifact store or reach a third party's production workers — while the accountable party never intended an attack. None of the six groups covers the accountable principal who lost scope control. That is an operator category, and it is a live question — for the overlay, and by the same high bar set out above. It is not a gap in the taxonomy, because the taxonomy never claimed that ground. R-SCOPE already classifies those steps correctly without knowing who the actor was, which is Axiom IV working exactly as designed.
And the argument here is falsifiable in the framework's own style. If someone produces an attack step whose classification or defence genuinely depends on the swarm being an actor rather than a configuration — a decision that a capability-vector reading cannot reach and an actor-group reading can — the position stated here is wrong and should change. The standing challenge sets out what such a case has to show.
Further Reading
- Twenty Agent Incidents, Ten Clusters — the classified corpus the vector above is derived from
- They Patched the Path. The Cluster Came Back in 24 Hours. — the takedown and re-formation, in detail
- Two Theses from the AI Hacking Incidents — the capability-vector position this post extends
- Attacker Profiling and the Actor Profile Designer — the six reference groups and the tool for scoring a vector
Sources and standing. Behavioural quotations and the July timeline are from OpenAI's published technical report. The per-cluster counts come from the twenty classified records in the register; the maturity scores are this post's interpretation of those counts and are offered to be argued with, not as measurement. No individual is named in the report and none is named here.