Connection Escrow proposes keeping an agent's network request from sending packets until a policy verifier approves the live socket. The prototype binds the decision to specific process/cgroup and destination details with a bounded expiry, then mediates it in the kernel before packets leave. This is narrower than allowing an agent to send arbitrary traffic…
Inspect the ClaimsCase · DiggingBeagle record
Connection Escrow tests signed agent network approval at the kernel socket boundary
A Linux eBPF prototype holds outbound connections until policy review authorizes the exact socket and destination.
Controlled Linux eBPF prototype with specified kernel and network tests in October 2026; not a production rollout.
Assessment profile
Assessment method- Setting
- research demonstration
- Exploitation
- demonstrated
- Evidence
- Grade B
- Remediation
- not applicable
- AI role
- AGAINST AI
Basis and provenance
Technical report or researcher-controlled experiments document the stated mechanism, but independent reproduction and population-wide prevalence are not established. Impact band is left unassigned because no calibrated damage threshold or representative victim impact is supported by the available record.
Assessed Oct 20, 2026 using diggingbeagle.assessment/1.
Evidence basis: CLM-ESCROW-EVAL · CLM-ESCROW-SOCK · CLM-ESCROW-VALIDATION · Connection Escrow: eBPF holds agent egress until policy verifier approval
At a glance
An enforceable denial below a model and its tools is more meaningful than a model statement that it will refrain from sending data.
Read the implicationsUDP and DNS bypass resistance, namespace changes, inherited connections and external policy verifier integrity require further testing.
Limits and uncertaintyFull account
Connection Escrow proposes keeping an agent's network request from sending packets until a policy verifier approves the live socket. The prototype binds the decision to specific process/cgroup and destination details with a bounded expiry, then mediates it in the kernel before packets leave. This is narrower than allowing an agent to send arbitrary traffic after a one-time broad approval. The paper reports a series of adversarial tests, concurrent connections and low added hold times in its configured Linux environment.
The mechanism is relevant to a malicious LoRA that leaked by DNS, but the paper does not demonstrate protection against every DNS transport, namespace transition, subprocess race or inherited long-lived socket. It is an experimental egress mediator, not independent proof that all generative agents are contained. The essential control point is below the language model, so a confident textual request cannot independently grant network access.
The proposal's distinctive control is a per-socket hold before traffic is sent. Authorization is tied to socket identity and destination, not just to the tool-call text that requested a connection. The authors evaluated kernel behavior and a separate policy verifier under defined Linux and Kubernetes environments; they did not establish that the verifier's classification of malicious content is always correct. Unsupported kernel versions and untested runtime paths remain separate from measured hold latency.
Timeline
- Oct 19, 2026disclosure
Public disclosure
Claims & evidence
CLM-ESCROW-EVALThe paper reports a limited adversarial network test suite and sub-10-ms median socket-hold observations, not universal isolation certification.supported
Basis: reported finding
- supportsConnection Escrow: eBPF holds agent egress until policy verifier approvalprimary disclosure
Evaluation section: adversarial cases and latency
CLM-ESCROW-SOCKThe authors describe binding outbound connection approval to the process context, destination and actual socket before packet transmission in a Linux prototype.supported
Basis: reported finding
- supportsConnection Escrow: eBPF holds agent egress until policy verifier approvalprimary disclosure
Mechanism and kernel hook design
CLM-ESCROW-VALIDATIONThe authors report a 36-check adversarial test suite and 600 socket-gated connections across six pods with a measured median hold of 4.3 ms; these measurements do not validate the correctness of every policy verdict.supported
Basis: reported finding
- supportsConnection Escrow: eBPF holds agent egress until policy verifier approvalprimary disclosure
Abstract, Evaluation: adversarial suite, Kubernetes cluster and released-connection latency
Implications
An enforceable denial below a model and its tools is more meaningful than a model statement that it will refrain from sending data.
Unknowns and contradictions
- UDP and DNS bypass resistance, namespace changes, inherited connections and external policy verifier integrity require further testing.
Sources and citation
Material revision history
- Oct 20, 2026 · Canonical change recorded · new in release · revision 177