Blog / AI Security

Two Theses from the AI Hacking Incidents

Agents act as coordinated attackers, and command and control moves inside the target. Both claims hold. Neither touches the ten clusters. Here is where each one lands in the framework.

Bernhard Kreinz • • Loading... • TLCTC v2.5.1
In one paragraph

The intrusions reported over the past year, in which an agentic coding tool ran most of an espionage campaign with a human only at the gates, and in which malware called a model at runtime to decide its next move, are read by many as a new kind of actor and a new kind of infrastructure. TLCTC reads them as two things it already has words for: a changed capability vector that compresses Δt and enables fan-out, and a boundary crossing that has turned around. The clusters, the axioms and the rules do not move.

Thesis A: Agents Are Coordinated Attackers

The claim is that agents must be positioned as threat actors in their own right, and that they may coordinate. The observation behind it is solid. One operator now runs many parallel workers, each one reconnoitring, exploiting, harvesting and moving on without waiting for a human, and the operator's role shrinks to approving a handful of decisions per target.

The taxonomy answers in two parts. First, the actor is still the party with intent. R-SCOPE opens with two questions: is there an actor, and did they intend the outcome. An agent has no intent of its own; the intent sits with whoever tasked it. When an agent inside a target follows injected instructions, the framework already says it plainly: the cluster was never wrong, the actor was. And Axiom IV forbids actor identity from structuring the clusters at all, which is why “AI agent” can no more become a classification input than “nation state” can.

Second, what actually changed is measurable elsewhere. The framework's actor groups carry a capability vector, and agentic tooling changes that vector along three axes the incidents make visible:

  • Velocity. Chains that used to run over days now run in minutes end to end. In TLCTC that is Axiom IX: every →[Δt=…] in the path shrinks, and whole chains drop from VC-1 or VC-2 into VC-3 and VC-4. The detection window that sits above each System Risk Event shrinks with them.
  • Fan-out. Coordination is a path shape, not an actor property. Genuinely concurrent steps are already (#X + #Y); many workers running the same chain against many targets is the replication pattern the ×N proposal describes. It stays a proposal, and these campaigns are the first real evidence it was worth writing.
  • Persistence without fatigue. The “APT in a box” pattern: a fast compromise followed by patient, low-noise operation, #7 →[instant] #4 →[instant] #1 and then VC-1 for as long as it takes.

So the defensible form of Thesis A is: agents are a capability vector that compresses Δt and enables fan-out, wielded by an actor who still owns the intent. “New threat actor” works in a briefing. Taken literally it breaks Axiom IV and invites an eleventh cluster, and the framework has already declined that invitation.

Thesis B: Command and Control Moves Inside

The second claim is the more interesting one. Classic command and control is an outbound beacon to attacker-owned infrastructure, which is exactly why egress monitoring, DNS analytics and proxy logs earn their keep. If the attacker instead steers an agent the organisation already runs, the channel turns around. Instructions arrive inbound, hidden in content the agent is built to ingest: a ticket, a document, a pull request, a web page. The agent's outbound traffic goes to a sanctioned model endpoint and to tools it is entitled to call. There is no attacker-owned egress to catch.

Classic C2Inverted C2
Directionoutbound beaconinbound instruction, outbound via sanctioned APIs
Infrastructureattacker-ownedthe target's own agent and its providers
What egress monitoring seesthe beaconnormal traffic to normal endpoints
Cluster of the delivery step#7 (foreign code executes)#1 (a designed ingestion function is abused)
Where the SRE sitsat executionat injection, before any traffic leaves
Boundary operator||[human][@External→@Org]|| on the way in||[api][@Attacker⇒@Source→@Org]||, a transit crossing

In notation the two shapes look like this. The classic chain:

#9 ||[human][@External→@Org]|| →[Δt=h] #7 →[Δt=d] #1 + [DRE: C]

The inverted chain, as the framework's indirect-prompt-injection path already writes it:

#1 ||[api][@Attacker⇒@Wiki→@Org]|| →[Δt=s] #1 →[Δt=s] #1 + [DRE: C]

Three rules do the work here, and none of them is new:

  • The transit operator, not #10. The wiki, the ticketing system or the model provider that carries the instruction is a transit party, ⇒. It becomes #10 only if the provider itself is compromised at a Trust Acceptance Event. And R-TRANSIT-3 still holds: the agent runtime on the target is attack surface, never transit.
  • R-SCOPE settles the entitlement question. The agent's credentials are legitimate. The actor steering it has no grant. Every action the agent takes for that actor is therefore Attack row, the same way stolen credentials are never in-grant. This blocks the tempting “but the agent was authorised” reading before it starts.
  • The SRE lands early. Loss of control over the agent, and over everything it can reach, happens at the injection step. That is where the Δt window opens, and it is well before the point where any egress control would see anything.

What Changes, and What Does Not

Nothing in the cluster layer. The incidents confirm the ten rather than strain them: the delivery is #1 or #9, execution is #7, credential use is #4, and the outcome is a Data Risk Event like any other. What changes is practical, and it changes in three places the framework already names.

  1. Actor groups gain a capability-vector note: agentic tooling compresses Δt, enables fan-out and removes fatigue. It does not add a group.
  2. Detection strategy moves up the chain. When the command channel is inbound and the egress is sanctioned, the compromise-detection window is the injection step, not the beacon. That is a Layer 2 question about which sources an agent is allowed to treat as instructions.
  3. The replication notation now has motivating incidents. It remains a proposal until the owner decides otherwise.

The short version. Agents are a new capability vector, not a new actor, and inverted C2 is a boundary that turned around, not a new cluster. The framework was built to absorb exactly this kind of change without moving its foundations.

Further Reading