Case · DiggingBeagle record

A composed AI-assisted review pipeline surfaced 66 previously unreported TCP/IP findings, but expert validation dominated final confidence

Forescout combined static query rules, source and binary analysis, LLM-assisted investigation and controlled testing across ten open-source and proprietary embedded networking stacks. The task was to prioritize and reproduce memory, integer and parsing errors before approaching vendors.

Public disclosure: Nov 12, 2026

Scope

Defenders inspect embedded TCP/IP parsers and protocol state machines with static rules and model-ranked leads.

UnratedAI role: WITH AIAssessment method

At a glance

Mechanism and trust boundary

  1. 01

    Starting condition

    Defenders inspect embedded TCP/IP parsers and protocol state machines with static rules and model-ranked leads.

    Boundary: precondition and attacker capability

  2. 02

    Trust boundary crossed or evaluated

    The report distinguishes 362 initial candidates, 312 after deduplication and 99 confirmed findings, including 33 previously known issues and 66 new findings. Fifty-four new issues were reported to maintainers. The model was one component of a workflow with substantial deterministic and human review.

    Boundary: mechanism

  3. 03

    Observed consequence

    An automated finding crosses from plausible pattern to credible security result only when the vulnerable path and corrective patch are reproduced and reviewed.

    Boundary: observed effect and limitations

Claims & evidence

CLM-NOV15-FORESCOUT-MECHForescout combined static query rules, source and binary analysis, LLM-assisted investigation and controlled testing across ten open-source and proprietary embedded networking stacks. The task was to prioritize and reproduce memory, integer and parsing errors before approaching vendors. The report distinguishes 362 initial candidates, 312 after deduplication and 99 confirmed findings, including 33 previously known issues and 66 new findings. Fifty-four new issues were reported to maintainers. The model was one component of a workflow with substantial deterministic and human review.supported

Basis: reported finding

Link to claim
CLM-NOV15-FORESCOUT-SCOPEA newly reported vulnerability is not evidence of exploitation; affected device versions, patch availability and exact CVEs vary. The published effort and estimated model costs do not include the cost of building the harness or all future remediation. The evidence is bounded by the cited source and test setting. Broader incidence, other environments, and actual production consequences are not established merely by these observations.supported

Basis: reported finding

Link to claim

Implications

An automated finding crosses from plausible pattern to credible security result only when the vulnerable path and corrective patch are reproduced and reviewed.

Controls and mitigations

  • Pair LLM-generated candidate explanations with executable regression and patch tests for each affected stack.
  • Track confirmed unique findings per human review hour and vendor-fixed release, not only raw scanner candidates or CVSS worst cases.

Unknowns and contradictions

  • The evidence is bounded by the cited source and test setting. Broader incidence, other environments, and actual production consequences are not established merely by these observations.

Sources and citation

Material revision history

  1. Nov 15, 2026 · Canonical change recorded · new in release · revision 250