Skip to content

Infinity-Context adversarial audit

Subject: v1.1.0 at a996d59f1eb468e7c386747e15f4c7e58d08410e. Branch: codex-audit, created from ci/release-workflow. Audit date: 2026-11-20 America/New_York / 2026-11-21 UTC.

This report suite implements the approved audit plan. It contains new findings, source checks, reproducible isolated experiments, primary-source research, prior-finding reassessment and remediation recommendations. It makes no production fixes. The starting AGENTS.md modification and prior audit/reference material are preserved. No commit, push, merge, release, paid provider call, live-agent configuration or user-memory-store change was made.

Start with the executive assessment, then the consolidated register and remediation roadmap.

Report Purpose
00 — Executive assessment Principal risks, strengths, readiness and validation limits
01 — AHTL conformance Complete PDF provenance/page inventory, normative matrix, codec/redaction findings
02 — Mathematics and retrieval Independent derivations, index snapshots, graph oracle and 27 dataset scaling matrix
03 — Integrity, persistence, workspace and daemon Exact bytes, ownership/seals, transactions, crash and failed-flush evidence
04 — Bus and concurrency Sequence/loss invariants, corruption/reset behavior, real separate-process propagation
05 — Integrations and proxy Claude/Codex/agy/Hermes/MCP contracts, HTTP/SSE controls and rejected curl lead
06 — Validation and performance Exact gates, coverage, fuzz/mutation outcomes, portable/native repetitions
07 — Supply chain and governance Dependencies, workflows, release controls, licenses and qualified legal questions
08 — Claims and documentation Material-claim matrix, current-vs-historical distinctions, Unicode table check
09 — Context budget and enhancements Unbounded epoch-window probe and measured/source-based optimization proposals
10 — Findings register / CSV Stable IDs with categories, severity and links to complete evidence
11 — Remediation roadmap Priority, dependencies, effort, compatibility and acceptance checks
12 — Research ledger Primary-source versions/access dates, 22 arXiv identity checks and research gaps
13 — Prior reassessment Every one of 51 actual old finding rows, preserving scoped judgments
14 — Coverage ledger All 519 tracked paths accounted for; review depth and explicit exclusions
15 — Rejected leads/procedures Counterevidence and reproducibility entry points

Evidence and method

Baseline, per-path inventory, claim leads, command/exit ledgers and sanitized harnesses live in evidence/. Mutable builds, mutations, synthetic keys/workspaces/databases and child-process experiments use /tmp/Infinity-Context/. Exceptions and environment failures are recorded in validation procedures. The reference PDF itself was not modified or duplicated.

Findings require a reachable source path or a reproducible contract counterexample and record prerequisites, impact, confidence, counterevidence, recommendations and acceptance criteria. Security conclusions state the relevant trust boundary; robustness/library defects are not automatically treated as remote exploits. Reserved IDs and rejected leads do not inflate counts. Independent root rechecks reproduced the CLI YAML leak, varied 10k recall result, daemon ACK/loss and failed-final-flush behavior. A countertest rejected the original curl-retention allegation.

Closeout checks are retained in verification and REUSE results. The original AGENTS.md EOF whitespace warning is preserved with its starting diff. No audited input was changed to make a check pass.

Scope of assurance

The path inventory is complete, while review depth varies. Large historical fixtures, generated assets and external archives have targeted/sampled dispositions rather than blanket semantic approval. The 58-page PDF was completely extracted, with selected equations visually checked. Material claims are grouped and adjudicated; the 826 extracted candidate lines are not 826 independent proofs. Full mutation-space execution, formal concurrency validation, every crash point, all external citations/licenses, unsupported platforms and live model/provider behavior are explicit gaps. See report 14 before interpreting this as exhaustive certification.

Baseline correctness gates and adversarial findings answer different questions. Passing tests, 95% line coverage, doctor or REUSE does not certify product claims, secret recall, strong test oracles, live integration or redistribution rights. Strict benchmark failures and fuzz crashes are reported without production remediation because fixes were outside the approved scope.