How to use Hermes to audit consensus validation
Hermes Kanban separated a consensus audit into evidence-bearing tasks. The developer verified a signature-order flaw and released corrected vote validation.
atto.cash
Suggested prompt
Suggested prompt written by Agents In Use, not the source author. Replace bracketed inputs and review tool permissions first.
Audit this node at a pinned commit in a read-only checkout. Create one Kanban card per concrete trust-boundary question with relevant files and an evidence directory. Separate discovery, validation and reproduction workers; verify inbound vote signatures before trusting representative identity or weight, and give confirmed findings to me for release decisions.
Reach people exploring AI agents
See placements →Evidence at a glance
Sourced account- Source author
- Felipe Rotilho
- Original account
- Read the source ↗
- Published
- Source link checked
- Reported result
- A human-verified vote-signature ordering flaw was fixed in node 1.33 and assigned CVE-2026-73855. Codex independently rediscovered it but missed other audit findings.
- Human involvement
- Rotilho verified the dangerous path himself, blocked the release, fixed validation and decided which findings and reproductions matched the live system.
A linked account supports attribution. It does not mean we independently reproduced its result.
Report an issue with this case
Open correction emailOpens your email app with the case link included. Nothing is sent automatically.
How the agent did it
- Rotilho replaced broad review lanes with pinned, evidence-bearing Hermes Kanban cards.
- Separate discovery and reproduction workers investigated the inbound-vote trust boundary.
- The developer verified the finding and shipped signature-first validation in node 1.33.
Felipe Rotilho used Hermes to audit Atto's live cryptocurrency node after broad Codex review lanes lost context and evidence. Hermes Kanban cards narrowed each worker's task to one question, a pinned commit, relevant files and an artifact directory.
Discovery, planning, validation and reproduction became separate stages. Workers had to create focused follow-up tasks rather than exhausting one conversation on an entire investigation. A dedicated audit profile distinguished confirmed vulnerabilities from hardening ideas and kept the repository read-only while evidence was being established.
One investigation found that inbound vote paths trusted a representative identity and weight before validating the vote signature. Rotilho initially doubted the critical report, then followed the path himself and confirmed it. He corrected the ordering and released node version 1.33; the issue became CVE-2026-73855.
A later Codex run independently rediscovered the same flaw using native subagents, but missed less severe findings from the structured audit. The result is a verified live-system correction with preserved audit boundaries, not proof of complete security. Rotilho still wanted a professional independent review and retained the judgment about what should block a release.
Prompt
Suggested prompt (written by Agents In Use, not the author's)
Source: Atto developer blog · Felipe Rotilho
Frequently asked questions
What did the faulty ordering trust too early?
The representative identity and voting weight claimed by an inbound vote.
Why use Kanban cards?
They constrained context and preserved commit-specific questions and evidence.
Did Codex find everything Hermes found?
No. It rediscovered the critical flaw but missed several less severe findings.
More details
About this account
Rotilho replaced broad review lanes with pinned, evidence-bearing Hermes Kanban cards. Separate discovery and reproduction workers investigated the inbound-vote trust boundary.
Original developer account opened. Fix shipped in node 1.33 as CVE-2026-73855. Later Codex rediscovery missed other findings; the author still wants an independent professional audit. Suggested prompt written by Agents In Use; no original task prompt was published in the verified account.
Credit: Felipe Rotilho — Atto developer blog
Source published
Added to Agents In Use
Source checked
Tools and topics
Hermes Agent · Kanban · Codex · Pinned commits · Reproduction artifacts
Consensus security · Evidence boundaries
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.