Skip to content

Thorn

LychD has spikes.

Thorn is the whitehat security casework Composition. A defender can bring an incident affecting the operator or a separately authorized assessment, including one on a system outside LychD. The private case keeps the observations, the defender's findings, and the receipts for any protective action or disclosure. It also names who owes the next verification or correction.

The intended deterrence is that targeting participating defenders can expose a campaign and prompt intervention. Useful protection and findings do not require identifying a human controller. Whether participation discourages attacks remains an unmeasured proposition.

thorn.defense, revision 1, is the accepted reference contract under Workflow. The case store and its Patterns await implementation. State of Work records delivery.

Purpose and boundary

Thorn keeps an investigation answerable: what may be examined, what the evidence supports, which response is justified, what actually happened, and which obligations remain.

Thorn owns Adjacent responsibility
Security case, authorized targets/methods, leads, findings, and competing explanations Logs, sensors, detectors, and analysis tools supply replaceable mechanisms and attributed observations.
Response decisions, scope, verification, and recovery follow-up The target system and its action adapter execute under current authority and retain their effect records.
Security verification of a suspected weakness or remediation Maintenance owners control code changes, migrations, activation, and operational recovery.
Recipient-specific reports, shared findings, challenges, and corrections Each peer controls its own work and disclosure; external recipients own their intervention.

Autopoiesis names intended verified self-repair and extension under operator authority. Routine database or workflow optimization, code maintenance, backups, and whole-system restoration belong to that wider concern under existing owners. They are outside Thorn's casework. A repair receipt can inform a security retest; a healthy service does not by itself prove containment or close a security finding.

Tool availability supplies no permission. An assessment pins targets, methods, possible effects, and stopping conditions. Receiving an attack supplies no right to enter suspected third-party systems, coerce someone, expose private identities publicly, or retaliate. Security and IAM retain enforcement and authority. Thorn introduces no general system administrator, scheduler, or self-authorizing agent.

Signet owns compute-node privacy qualification and the certificate lifecycle. Thorn can supply an admitted security finding or repair retest for Signet to judge.

Typed findings, requests, and settled receipts can cross these boundaries without transferring ownership. A promise to coordinate separate Compositions' live Invocations needs the explicit Suite contract.

From encounter to result

A defender receives a suspicious message or an alert about unexpected account activity. Thorn preserves permitted evidence, checks the signal's freshness and scope, and opens or links a case. A human can instead begin with an authorized assessment question. Triage may also record a non-incident, a gap, or a concrete refusal.

The case proceeds through five decisions:

  1. What was observed, what is exposed, and what work is authorized?
  2. Which immediate protective action is admitted, and how will its effect be verified?
  3. Which optional lead could establish the next useful fact?
  4. Do the findings justify a particular report or should the case abstain?
  5. What needs correction, recovery, or further observation before the work can close?

Protection and investigation can proceed independently. Existing controls continue enforcing their rules when Thorn, a model, or peers are unavailable. Immediate protection does not wait for attribution or a completed inquiry. Detector interpretations remain distinct from observations and case judgments. Duplicate or stale alerts confer no new authority; missing capture or a failed sensor leaves a coverage gap rather than evidence of safety.

Follow the campaign develops the lead loop, peer sightings, supported activity paths, and handling of hostile evidence. Response policy independently sets effort, automation, participation, attention, and cost. Iron Shield favors quiet protection; Guided Inquiry proposes leads; Hornet pursues admitted leads persistently. Their limits do not alter authority or the evidence needed for a claim.

The case and its Patterns

The private case must explain both the investigation and its consequences. Retain its owner, restricted subject/victim references, purpose, scope, effective policy revision and expiry, budget allocations/consumption, leads, claims, contrary evidence, action attempts and receipts, dispositions, and correction obligations.

For a reviewer to challenge a claim, its source references must preserve custody, method, transformations and redaction lineage, original and UTC times with clock uncertainty, and observed versus inferred status. Original bytes and local custody digests remain with their owners. Each assessment records its reviewer, grounds for confidence, and review/expiry conditions.

Pattern Admitted input → accounted result
thorn.open_case@1 Scoped report, observation, or assessment question → triage and a new/linked private case, non-incident disposition, or refusal.
thorn.investigate@1 Admitted leads or assessment and source policy → supported claims, competing explanations, gaps, and disposition of each lead.
thorn.protect@1 Supported response and current authority → proposal or admitted action attempt, verified/unknown result, and recovery obligations.
thorn.report@1 Findings and recipient policy → exact reviewed packet and separately authorized handoff disposition.
thorn.correct@1 Changed or challenged evidence → successor claims, invalidated derivatives, and recipient-specific correction dispositions.

An operator can admit protection alone, investigate without reporting, or reopen a finding for correction after the investigation ends. The case owner sees outstanding effects and obligations beside each finding. A peer sees only the admitted exchange. An external recipient sees the exact packet approved for that destination.

For example, a defender reviewing a synthetic sign-in alert may inspect one account's audit records for the reported interval. Two alerts cite the same denied sign-in. The case links both to that observation. The reviewer records one denied attempt, a gap in coverage for subsequent sign-ins, and an unknown actor. With no authority to contact peers or change the account, the owner can close triage with those limits and abstain from reporting. If a corrected audit record later changes the finding, the case retains the original judgment and its successor.

Respond and verify

Before attempting a response, bind it to the target, evidence conditions and freshness, current authority, impact, duration, limits, verification, and recovery plan. The effect boundary checks that account independently of the worker proposing it. It must also prevent concurrent requests from duplicating or conflicting over an effect. Assign an accountable owner who retains verification and recovery even if the worker disappears.

Keep proposed, authorized, attempted, partially effective, maintained, and recovered states distinct. An accepted command does not establish containment. Verification may reveal a failed action, missing visibility, or an unknown effect. Reconcile uncertainty before retry so another attempt cannot silently duplicate a change.

Account for false positives and service impact. Rollback or expiry requires the declared recovery decision against current exposure; blindly restoring a still-compromised capability is not recovery. An irreversible action needs an admitted recovery path. A maintenance outcome and the security retest of that outcome remain distinct findings.

Reports and corrections

A report separates facts from inference and retains contrary evidence, custody, gaps, affected scope, and the requested response. Review whether the evidence supports that intervention and its impact on unrelated users of shared infrastructure. Human acceptance records a decision; it cannot make an allegation true. Narrow an unsupported request or abstain.

Independently verify the recipient and destination, accepted threat category, purpose, and mandate/jurisdiction where applicable. A human accepts the exact packet, destination, and purpose; changing them requires renewed acceptance. Preparation does not authorize sending or publication. Outgoing evidence uses only recipient-permitted opaque references or scoped EvidenceDigest@1 under Security's disclosure rules, not raw canonical source digests. Peer exchange follows the same disclosure boundary through its own contract.

Once a report is admitted for handoff, persist its exact artifact, authority, attempt identity, and receipts. Record whether it was prepared, sent, acknowledged, refused, failed, or remains unknown. A successful transport proves neither guilt nor official action. If the send may have happened but its acknowledgement was lost, reconcile that attempt before retrying.

Review equivalent allegations across visible collaborating cases, including differently worded reports, while recording visibility limits. Cooperation must not become report flooding. A correction is a separately admitted successor with its own recipient disposition, not another copy of the allegation. Findings can remain useful when external reporting is unwarranted.

Finish, stop, or return

Containment, operational recovery, investigation completion, and report delivery have separate dispositions. Actor unknown and insufficient evidence are valid investigative findings. Partial completion names remaining obligations and owners; refusal names the unavailable scope or boundary; failure names the failed operation. Unknown effects keep overall settlement unresolved even when analysis has finished.

Cancellation or revocation stops newly disallowed work and disclosure and initiates authorized collection shutdown. It cannot undo delivered effects. Resume rechecks current permissions, evidence freshness, sensor generation, and outstanding attempts. A stopped inquiry cannot erase an unfinished action, delivery, or recovery obligation.

Corrections can outlive a closed case. Successor claims update dependent findings and affected recipients; unresolved acknowledgements remain visible. Retention/deletion follows case and custody policy while preserving only permitted lineage. Remote deletion or recall is not guaranteed.

First proof

Thorn Trials begins with a synthetic local attack and simulated response, then compares the same case across three simulated defenders. A scoped assessment and repair retest separately exercise security verification. An implementation must improve a supported decision or reduce total work against competent simpler protection, within authority, privacy, recovery, and cost limits. The trials page defines defeat conditions and advancement evidence; no trial result is claimed.

Composition Portfolio