---
title: 'Clawgang: Graph Theory & AI Agent Security'
url: https://www.emergentmind.com/topics/clawgang
type: topic
---

# Clawgang: Graph Theory & AI Agent Security

“Clawgang” is an *Editor’s term* for a body of research organized around two distinct uses of “claw.” In graph theory, the claw is the star \(K_{1,3}\), which appears as a forbidden induced subgraph, as the target of deletion and online Ramsey problems, and as a structural constraint on power graphs of finite groups. In recent AI-systems security work, “Claw-like” denotes always-on personal AI agent processes with persistent access to files, shell, credentials, tools, and external services; the associated literature studies both their architectural attack surfaces and out-of-band workflow verification [1309.1835] [2606.30755] [2605.06205]. The grouping is editorial rather than standard, but it captures a coherent research pattern: a simple local obstruction or privileged runtime feature induces strong global structure.

## 1. Claws, co-claws, and the structural theory of claw-free graphs

A graph is claw-free if it has no induced copy of \(K_{1,3}\), where \(K_{1,3}\) is the claw: one center vertex adjacent to three pairwise nonadjacent vertices. The complementary obstruction is the co-claw \(\overline{K_{1,3}}\). The class
\[
\mathrm{Forb}\{K_{1,3},\overline{K_{1,3}}\}
\]
therefore consists of graphs with no induced claw and no induced co-claw. Its main structural characterization is exact: a graph \(G\) belongs to this class if and only if \(G\) is one of the following: the exceptional graph \(A_6\); an induced subgraph of the Paley graph \(P_9\); a graph whose connected components are cycles of length at least \(4\) or paths; or the complement of one of the graphs in the previous two categories [1309.1835].

The exceptional configurations are sharply delimited. The Paley graph \(P_9\) is self-complementary, satisfying
\[
P_9 \cong \overline{P_9},
\]
and \(A_6\) together with \(\overline{A_6}\) is singled out as especially important because these are the only graphs in the class that contain both a triangle and an independent set of size \(3\) with no common vertex.

A classical bridge used in this analysis is line-graph theory via the edge-graph \(S(U)\). Its vertices are the edges of a graph \(U\), and two edges \(xy\) and \(xz\) are adjacent in \(S(U)\) when \(yz\) is not an edge of \(U\). The key equivalence is
\[
U\in \mathrm{Forb}\{K_{1,3}\}\quad \Longleftrightarrow\quad S(U)\in \mathrm{Forb}\{K_3\}.
\]
Thus claw-freeness in \(U\) is translated into triangle-freeness in \(S(U)\). A plausible implication is that the local induced-star obstruction is unusually amenable to transfer into secondary graph constructions, which helps explain why claw-free phenomena recur across otherwise different combinatorial problems.

## 2. Homogeneous triples, Boolean sums, and reconstruction up to complementation

For a graph \(G\), a subset of vertices is homogeneous if it is either a clique or an independent set. The \(3\)-uniform hypergraph
\[
\mathcal H^{(3)}(G)
\]
has as hyperedges exactly the \(3\)-element subsets of \(V(G)\) that are homogeneous in \(G\), that is, those inducing either \(K_3\) or \(\overline{K_3}\). The associated reconstruction problem asks which graphs are determined, up to complementation, by these \(3\)-homogeneous triples [1309.1835].

The technical mechanism is the Boolean sum of two graphs \(G\) and \(G'\) on the same vertex set:
\[
G\dot{+}G',
\qquad
E(G\dot{+}G')=E(G)\Delta E(G').
\]
The central equivalence states that, for a graph \(U\), the following are equivalent: there exist graphs \(G,G'\) with the same \(3\)-homogeneous subsets such that \(U=G\dot{+}G'\); both \(S(U)\) and \(S(\overline U)\) are bipartite; and either \(U\) is an induced subgraph of \(P_9\), or the connected components of \(U\), or of \(\overline U\), are cycles of even length or paths.

A key lemma makes the link explicit. If \(U=G\dot{+}G'\), then equality of the \(3\)-element homogeneous subsets of \(G\) and \(G'\) is equivalent to an edge-partition condition in the edge-graphs \(S(U)\) and \(S(\overline U)\). Concretely, with
\[
A_1=E(U)\cap E(G),\qquad A_2=E(U)\setminus E(G),
\]
the sets \(A_1,A_2\) form a bipartition of \(S(U)\) into independent sets, and an analogous statement holds for \(\overline U\).

The role in reconstruction is direct. Two graphs are isomorphic up to complementation if one is isomorphic to the other or to its complement; they are \(k\)-hypomorphic up to complementation if every \(k\)-vertex induced subgraph of one is isomorphic to the corresponding subgraph of the other or to its complement; and a graph is \(k\)-reconstructible up to complementation if this local condition forces global isomorphism up to complementation. The Boolean-sum characterization constrains the nontrivial possibilities for \(U=G\dot{+}G'\) to paths, even cycles, induced subgraphs of \(P_9\), and complements of these. This suggests that the hypergraph of homogeneous triples functions as a highly restrictive invariant rather than a merely coarse summary.

## 3. Claw deletion and approximation on split and bipartite graphs

For \(t\ge 3\), \(K_{1,t}\) is called a \(t\)-claw, and the minimum \(t\)-claw deletion problem (\texttt{Min-\(t\)-Claw-Del}) asks for a minimum-size vertex set \(S\) such that \(G[V\setminus S]\) is \(t\)-claw free. In a split graph, the vertex set can be partitioned into a clique and an independent set, and every \(t\)-claw has its center vertex in the clique partition. This one-sidedness motivates the minimum one-sided bipartite \(t\)-claw deletion problem (\texttt{Min-\(t\)-OSBCD}): given a bipartite graph \(G=(A\cup B,E)\), find a minimum vertex set \(S\) such that \(G[(A\cup B)\setminus S]\) has no \(t\)-claw with the center vertex in \(A\) [2306.13306].

The approximation landscape is tight. A primal-dual algorithm approximates \texttt{Min-\(t\)-OSBCD} within a factor of \(t\), and it is \(\UGC\)-hard to approximate with a factor better than \(t\). The paper also gives a dense-instance improvement: if \(d(v)\ge 2(t-1)\) for all \(v\in A\), then the algorithm approximates \texttt{Min-\(t\)-OSBCD} within a factor of \(2\). The submodular formulation uses
\[
f_t(F)=2\left[\sum_{v\in A}\min\{(t-1),d_F(v)\}\right],
\]
with the statement that \((E,f_t)\) is a \(2\)-polymatroid and that \(F\subseteq E\) is a matching in \((E,f_t)\) if and only if the corresponding subgraph is one-sided \(t\)-claw free.

The split-graph problem inherits hardness by a direct construction: from a bipartite instance \(G=(A\cup B,E)\), form a split graph \(H=(A\cup B,E')\) with
\[
E'=E\cup \{(u,v)\mid u,v\in A\}.
\]
Then \(S\) is a one-sided \(t\)-claw deletion set in \(G\) if and only if \(S\) is a split-\(t\)-claw deletion set in \(H\). Consequently, assuming \(\UGC\), \texttt{Min-\(t\)-Claw-Del} on split graphs cannot be approximated better than \(t\), and the paper also states that \texttt{Min-\(t\)-Claw-Del} is \(\NP\)-complete even when each vertex \(v\in A\) has at least \(2(t-1)\) neighbors in \(B\).

The complementary maximization problems, Max-\(t\)-OSBC-Subgraph and Max-\(t\)-Claw-Subgraph, are treated as the natural dual formulations. They are \(\APX\)-complete and admit \(2-\frac{1}{t}\) approximation algorithms. Within this line of work, claw-freeness is not merely a hereditary graph property; it is the defining feasibility condition for a family of tight approximation thresholds.

## 4. Claw-free reduced power graphs of finite groups

For a finite group \(G\), the undirected power graph \(P(G)\) has vertex set \(G\), and two distinct vertices \(u,v\) are adjacent if one of the cyclic subgroups \(\langle u\rangle,\langle v\rangle\) is contained in the other. The reduced power graph \(P^*(G)\) is the induced subgraph of \(P(G)\) on \(G\setminus\{1\}\). The classification problem asks which finite groups have claw-free \(P^*(G)\) [2407.20110].

A first structural fact is that any claw in \(P^*(G)\) must appear in one of two orientations: either \((b\to a_1,a_2,a_3)\) or \((a_1,a_2,a_3\to b)\). This immediately constrains element orders. If \(g\in G\) has order divisible by three distinct primes \(p_1,p_2,p_3\), then
\[
(g\to g^{p_1},g^{p_2},g^{p_3})
\]
is a claw of the first type. If \(g\) has order \(p_1p_2\) with \(p_1\neq p_2\), then
\[
(g\to g^{p_1},g^{p_2},g^{p_1p_2})
\]
is also a claw of the first type.

The nilpotent case is highly restrictive. For odd \(p\), a finite \(p\)-group has claw-free \(P^*(G)\) if and only if it is cyclic or \(\exp(G)=p\). For noncyclic \(2\)-groups, claw-freeness forces either \(\exp(G)\le 4\) or \(G\) to be dihedral. More generally, if \(G\) is nilpotent and \(P^*(G)\) is claw-free, then either \(G\) is a \(p\)-group, or \(G\) is cyclic; in the cyclic non-\(p\)-group case, \(|G|=p^mq\) for distinct primes \(p,q\).

Beyond nilpotent groups, the paper proves that if \(P^*(G)\) is claw-free, then either \(G\) is solvable or \(G\) is almost simple; in the latter case, the socle of \(G\) is isomorphic to \(PSL(2,q)\) for suitable choices of \(q\). For finite non-abelian simple groups, the classification is exact: \(P^*(G)\) is claw-free if and only if \(G\cong PSL(2,q)\) and, writing \(d=(q-1,2)\),
\[
\frac{q-1}{d},\qquad \frac{q+1}{d}
\]
must each be either a prime power or the product of a prime and a prime power. The global numerical corollary is that if \(P^*(G)\) is claw-free, then the order of \(G\) is divisible by at most \(5\) different primes.

This program shows that a graph-theoretic prohibition in \(P^*(G)\) constrains centralizers, Sylow structure, solvability, and the simple-group composition factors. A plausible implication is that claw-freeness acts here as a low-complexity certificate for rank-one-like subgroup geometry, with \(PSL(2,q)\) emerging as the surviving simple family.

## 5. The claw in online Ramsey theory

The online Ramsey number \(\tilde r(G,H)\) is defined through the Builder–Painter game on an empty graph with countably many vertices. In each round, Builder reveals an edge, Painter colors it red or blue, and Builder wins once a red copy of \(G\) or a blue copy of \(H\) appears. For the claw versus cycles, the exact result is
\[
\tilde r(K_{1,3},C_\ell)=\left\lfloor \frac{3(\ell+1)}{2}\right\rfloor
\qquad\text{for all }\ell\ge 13
\]
[2601.05452].

The lower bound uses a Painter strategy that colors an exposed edge red unless doing so would create either a red \(K_{1,3}\) or a red cycle \(C_k\) with
\[
3\le k\le \left\lfloor\frac{\ell+1}{2}\right\rfloor.
\]
This keeps the red graph claw-free, of maximum degree at most \(2\), and without short cycles, so it is a disjoint union of paths. If \(s\) is the number of red path components and \(X\) the set of vertices of degree \(2\) in the red graph, the paper derives
\[
|X|+s\le \left\lfloor\frac{\ell+1}{2}\right\rfloor,
\]
while a blue \(C_\ell\) would require
\[
2|X|+s\ge \ell.
\]
The contradiction yields the lower bound.

The upper bound is constructive. It introduces good, better, and best blue paths according to the number of incident red edges at their endpoints; three local gadgets called Type 1, Type 2, and Type 3; and the wavy path \(P(i,j)\), consisting of a blue path \(P_{i+j}\) together with \(j\) pairwise vertex-disjoint \(P_3\) subpaths whose endpoints are joined by red edges. Builder forces one of four small initial configurations \(G_1,G_2,G_3,G_4\), then applies extension lemmas and a Blue Path Expansion Algorithm until a blue \(C_\ell\) is forced within the exact budget.

This result sits within a specific prior literature. Cyman, Dzido, Lapinskas, and Lo determined exact values for \(G=P_3\), Song, Wang, and Zhang proved
\[
\tilde r(K_{1,3},P_\ell)=\left\lfloor \frac{3\ell}{2}\right\rfloor,
\]
and Adamski, Bednarska-Bzdęga, and Błażej obtained asymptotically tight results when \(G\) is an even cycle. Against that background, the claw-versus-cycle theorem gives an exact long-cycle value for a sparse forbidden red graph.

## 6. Claw-like agents as computer systems

In AI-systems security, Claw-like agents are always-on personal AI agent processes running inside the user’s own computing environment with persistent access to high-value local resources: files, shell, credentials, tools, and external services. OpenClaw is used as the exemplar. These agents install packages, keep state across sessions, schedule subtasks, and mediate I/O. The central analytical move is to treat such an agent as an agentic computer system: the gateway runtime is like an OS/runtime layer, Skills are like user-installed applications, and Plugins are like in-process loadable extensions [2606.30755].

The paper maps agent components to classical system components and corresponding protection gaps. A package repository corresponds to a Skills marketplace; process address space corresponds to the LLM context window; file system or storage corresponds to persistent memory; user-installed applications correspond to Skills; IPC channels correspond to connectors such as email or Slack; in-process loadable extensions correspond to Plugins; the audit subsystem corresponds to gateway logs; and user input versus data plane corresponds to document or email content inside context. The associated protection mechanisms—code signing, review, sandboxing, isolation, DAC or MAC, authenticated channels, privilege separation, redaction, integrity protection, and data or instruction separation—are described as absent or incomplete on the agent side.

Five security principles are extracted: process isolation, least privilege, persistent-state protection, cross-boundary mediation, and data-instruction separation. These are grouped into four benchmark attack surfaces: Skill Supply-Chain Integrity (SSI), Persistent State Exploitation (PSE), Cross-Boundary Data Flow (CDF), and Indirect Prompt Injection (IPI). SafeClawArena operationalizes this framework as a benchmark of \(406\) adversarial tasks across these four attack surfaces, evaluated on three platforms—OpenClaw, NemoClaw, and SeClaw—and five frontier LLMs—GPT-5.1-Codex, GPT-5.4, Gemini-3-Flash-Preview, Gemini-3.1-Pro-Preview, and Claude-Opus-4.6.

The execution model is systems-oriented rather than prompt-only. Each task runs in a fresh Docker container reproducing a production deployment, with the gateway daemon, the LLM backend, the Sim-Google CLI, task-specific Skills, Plugins, or content, and seeded workspace files with credentials. The benchmark uses canary-marked credentials of the form
\[
\text{CANARY}_{\{KIND\}_{\{SUFFIX\}}},
\]
with suffix space
\[
36^6\approx 2.2\times 10^9,
\]
and deterministic taint matching across nine output channels: agent response, outbound message, local Sim-Google call log, memory write, gateway log, configuration write, workspace file write, webhook payload, and cron output.

The reported security picture is severe. Across all \(15\) configurations, overall attack success ranges from about \(20\%\) to \(70\%\); the lowest overall attack success is \(20.2\%\) for NemoClaw + Claude-Opus-4.6, while the highest is \(69.7\%\) for both OpenClaw + GPT-5.4 and NemoClaw + GPT-5.4. Category 1.4 Malicious Plugin succeeds in \(100\%\) of cases on every unhardened configuration regardless of LLM, because the Plugin runs as native code inside the gateway and bypasses the LLM entirely. SeClaw reduces GPT-5.4’s attack success rate from \(69.7\%\) on OpenClaw to \(21.9\%\), but some of that gain is attributed to removing attack surface rather than to active defenses. The paper’s broader claim is that Claw-like agents should be analyzed like operating systems, not just like language models.

## 7. Physical workflow auditing with ClawGuard

ClawGuard addresses workflow hijacking in autonomous LLM agents. In this setting, a user request is executed as an ordered sequence of skills or tool invocations, and an attacker may alter that sequence by inserting, omitting, reordering, substituting, or branching skills while keeping the overall conversation semantically plausible. The key claim is that workflow hijacking is also a physical execution problem: changes in the workflow should change the host’s hardware activity and hence its electromagnetic emanations, making passive out-of-band detection possible [2605.06205].

The intended workflow is modeled as
\[
W=\langle s_1,s_2,\dots,s_n\rangle,\quad s_i\in\mathcal S,
\]
and the compromised execution as
\[
W^\star=\langle s_1^\star,s_2^\star,\dots,s_m^\star\rangle.
\]
External receivers observe passive EM traces
\[
X=\{X_r(t;f_r)\}_{r=1}^R,
\]
with
\[
X_r(t;f_c)=H_r(f_c)\Big(\sum_{j\in\mathcal C}E_j(t;f_c)\Big)+N_r(t).
\]
ClawGuard implements a dual-SDR, drift-aware coarse–fine pipeline. Two HackRF One SDRs sample at \(20\,\text{MS/s}\) with \(8\)-bit IQ, placed about \(2\) cm from the host chassis. Carrier selection is empirical rather than assumed from hardware specifications; a survey over \(1\)–\(3000\,\text{MHz}\) produced the pair \((80,800)\,\text{MHz}\) for the replication corpus.

Signal processing uses coarse skill intervals \([a_i,b_i]\) and overlapping fine windows
\[
F_{i,j}=[a_i+j\rho,\;a_i+j\rho+\tau].
\]
Each fine window is converted into a \(320\)-dimensional V10 feature vector \(\mathbf{x}_{i,j}\in\mathbb R^{320}\), including spectral energy or log-PSD band energies, spectral shape statistics, temporal envelope features, and cross-receiver coupling features. Drift compensation applies cycle-local normalization,
\[
\tilde{x}_{c,k}=\frac{x_{c,k}-\mu_{c,k}}{\sigma_{c,k}+\epsilon},
\]
temperature detrending,
\[
x'_k=\tilde{x}_k-\sum_{\ell=0}^{d}\beta_{\ell,k}T^\ell,
\]
with \(d=1\), and ANOVA feature selection with top-\(k\) retention, using \(k=65\) in the focused three-skill task and \(k=80\) in larger settings. Because many features are more predictive of collection cycle than of skill identity on the big48 benchmark, evaluation uses leave-one-cycle-out cross-validation rather than random splits.

The system then separates skill evidence from attack detection. Stage 1 performs skill recovery on physically separable subsets via
\[
p_{i,j}(s)=\Pr(s\mid \mathbf{x}'_{i,j}),\qquad
\bar p_i(s)=\mathrm{Pool}_j\, p_{i,j}(s),\qquad
\hat s_i=\arg\max_s \bar p_i(s).
\]
Stage 2 performs attack-state detection using
\[
a_{i,j}=h_\psi(\mathbf{x}'_{i,j})\in\{\text{background},\text{normal},\text{attack}\},
\]
and a record is flagged hijacked when aggregation over fine-window states exceeds a threshold.

The experimental scale is unusually large for RF side-channel work: a \(7.82\) TB main corpus with \(12{,}232\) records across \(16\) benign skills and \(22\) attack skills, plus a separate new-bands replication corpus using \((80,800)\,\text{MHz}\). On a production split of \(11{,}800\) records—\(1{,}650\) attack and \(10{,}150\) normal—ClawGuard achieves \(\text{ROC-AUC}=0.9945\) and \(\text{PR-AUC}=0.9305\), with \(100\%\) true-positive rate at \(1.16\%\) false-positive rate. On the new-bands corpus, the same pipeline reaches \(83.6\%\) sub-window accuracy, \(88.3\%\) record-vote accuracy, and \(90.3\%\) attack recall. The paper also reports median inference around \(18\) ms.

The limitations are explicit. ClawGuard is not a universal \(22\)-class workflow recognizer, broad open-set classification is fragile, anomaly detectors such as IsolationForest, OneClassSVM, and Mahalanobis scoring are near chance, and generalization across devices and across days still needs recalibration. The stated contribution is narrower: an out-of-band, forge-resistant physical check against compromised host software, under a threat model in which the host OS may be fully compromised while the SDRs and policy channel remain trusted.

Source: https://www.emergentmind.com/topics/clawgang