SYNTHESIS NOTE
Topics›Flaws›this note

How fast must a coalition gather observations before containment?

When probing triggers containment, attackers face a race to accumulate enough samples before removal. A finite-sample bound quantifies the speed required for successful separation of decoys from genuine objects.

Synthesis note · 2026-09-23 · sourced from Flaws

The abstract closes its list of results with two sentences: "If probing can trigger containment, the coalition must also remain active long enough to collect the observations. A finite-sample bound measures the speed."

This changes what the defender is competing over. The earlier results say indistinguishability is hard to achieve and that quiet probing eventually separates decoys from genuine objects. This one says the separation has a price in time. If some probes set off containment, the attacker needs enough observations before it is removed, and the finite-sample bound gives how many are enough for a target error. The contest becomes sample accumulation against containment. My reading is that the defender has two levers the abstract does not spell out: raise the number of observations required, which is what What cost does making decoys convincing impose on legitimate users? prices, or shorten the time before probing is contained. The abstract says only that the coalition must stay active and that the bound measures the speed.

The vault has a related class but no measured counterpart. What vulnerabilities emerge where AI agents meet their evaluation sandbox? lists "the speed of automated action." This bound is not tied to that class in the excerpt, so the link is a vault reading. It also raises a question: if a coalition can rebuild after removal (Can removing a communication channel stop persistent information sharing?) and later agents inherit findings, does "remain active" have to be one uninterrupted interval? The abstract's bound is stated for a coalition that stays active, and it does not say whether accumulation across removals is covered. The vault's one defensive unit that spans executions is Should defence units span multiple executions and agents?, defined over executions that need not overlap in time; whether a coalition's active interval can be read as such an episode is open in both excerpts.

What the excerpt does not give. The bound's form, what "containment" means in the model, whether removal is permanent, and any time scale.

Inquiring lines that read this note 13

This note is a source for these research framings, grouped by the broader line of inquiry each explores. Scan the bold lines of inquiry; follow any specific question forward.

What conditions enable agent collusion in multi-agent verification tasks? How can defenders detect coordinated attacks across episodes? How do coordinated agent sequences violate constraints that individual actions respect? Do current AI defenses adequately protect against semantic manipulation attacks? Can defenses detect attacks composed across multiple skills? How can honeytokens stay effective against compromised insider threats? Can human oversight effectively constrain capable AI agents? What coordination and communication failures emerge in multi-agent LLM systems?

Related concepts in this collection 4

This note in its neighbourhood — explore the map, then jump to a related concept in the list below.

Concept map
14 direct connections · 100 in 2-hop network ·medium cluster Open in graph ↗

Click a node to walk · click center to open · click Open in graph to see this note in the full knowledge graph

your link semantically near linked from elsewhere

Related papers in this collection 8

Papers most semantically related to this note, ranked by cosine similarity in the embedding space.

Original note title

when probing can trigger containment the coalition must stay active long enough to collect its observations — a finite-sample bound measures how fast