Skip to content

Infinity-Context (icx) v1.0.0 — Adversarial Audit

  • Auditor: Claude (Opus 5.5) in Claude Code, at the maintainer's request.
  • Date: 2026-11-18.
  • Subject: branch claude-audit at c176d03 (v1.0.0). The audit changed no production code.

Start with 00-executive-summary.md.

Documents

Doc Workstream
00-executive-summary.md Top risks, claim-vs-reality scorecard, prioritized remediation
01-ahtl-spec-conformance.md W1: AHTL PDF vs Markdown vs icx-codec
02-integrity-and-lossless-claims.md W2: CAS, DAG, persistence, "lossless"
03-hdc-fractal-math.md W3: HDC, Fractal IFS, surprise gate, recall
04-bus-and-concurrency.md W4: seqlock telepathy bus
05-security-findings.md W5: threat model vs code, security defects
06-agent-integration-verification.md W6: Claude Code, Codex, agy, Hermes, MCP, provider caching
07-performance-and-benchmarks.md W7: SLO suite, real hook latency, profiling
08-test-quality.md W8: coverage, mutation testing, missing test classes
09-docs-citations-governance.md W9: cross-doc consistency, citations, licensing
10-ci-supply-chain.md W10: workflows, pinning, dependencies
11-enhancements-and-optimizations.md Forward-looking improvements
12-v1.1.0-remediation-status.md Per-finding status after the v1.1.0 release work
findings.csv Every finding: ID, severity, location, evidence type (ticket-ready)

Method

  1. Baseline. Before any analysis I ran every gate CI runs, exactly as CI runs it.
  2. Claim extraction. I listed every quantitative and normative claim across docs/, docs/ahtl/ (including the 58-page PDF, extracted with pdftotext), ref-docs/, README, CHANGELOG and ROADMAP.
  3. Verification. Each claim was checked against the code at a cited line. Wherever possible it was also checked against a reproduction: a PoC test or binary run against the workspace crates, a measurement with the release binary, a perf profile, or a mutation run.
  4. Upstream checks. External claims (agent hook APIs, MCP, provider prompt caching, arXiv citations, licenses) were checked against current upstream documentation and the locally installed agent binaries. Research subagents did that work, and I re-checked the highest-impact claims against primary sources myself.
  5. Severity.
Level Meaning
Critical Breaks the product's core promise in normal use
High A security defect with a realistic path, or a headline claim that is false
Medium A real defect or misstatement with bounded impact
Low Hardening, latent defects, minor inaccuracies
Info Dead code or a note
  1. Leads rejected. Pre-audit leads that failed verification are recorded as corrections, not findings:
  2. the slice-hint packet-level round trip actually works;
  3. the cache-hit filter does not exclude seal turns;
  4. the short-handle attack costs 2⁴⁸ work, not 2²⁴;
  5. ya29. and glpat- tokens are redacted;
  6. the arbitrary transcriptPath read cannot disclose content.

Disclosure. Security findings are described here at the level of defect, impact and fix. Exploit-level reproductions are kept outside this public repository (see 05-security-findings.md).

Baseline at c176d03

  • Toolchain: rustc 1.100.0. The host builds with -C target-cpu=native (from the global cargo config).
Gate Result
cargo fmt --all -- --check pass
cargo clippy --workspace --all-targets -- -D warnings pass
cargo test --workspace pass (45 tests)
RUSTDOCFLAGS=-D warnings cargo doc --workspace --no-deps pass
cargo llvm-cov … --fail-under-lines 95 pass: 95.75% lines; production-only 95.78%
icx benchmark × 10 all PASS (medians in 07-performance-and-benchmarks.md)
icx doctor READY (informational probes only)
cargo audit, osv-scanner 0 advisories (30 crates)
cargo deny check licenses FAIL: no deny.toml exists
uvx reuse lint pass (468 of 468 files)

Tool versions

Tool Version
cargo-llvm-cov installed
cargo-mutants 27.1.0
cargo-geiger 0.14.0
cargo-semver-checks 0.51.0 (n/a: unpublished)
cargo-msrv 0.19.6 (n/a: MSRV = stable)
cargo-machete 0.10.1
cargo-audit, cargo-deny, cargo-nextest, cargo-outdated installed
osv-scanner 2.7.0
semgrep 1.186.0 (p/rust, p/secrets)
gitleaks (CI pins 8.30.1)
typos-cli 1.52.2
lychee 0.24.2
markdownlint-cli2 installed
perf installed
pdftotext installed

Agent versions used for the integration checks: Claude Code 2.1.328, codex-cli 0.170.0, agy 1.5.11, Hermes source dc394c6.

Not done, and why

  • No live provider API calls. Real cache-read tokens were not measured; that needs API keys and a recorded session (ICX-AUD-002, -003).
  • No live agent sessions. Hook-output effects were verified against documentation and code, not by running the agents. Items that need a live test are marked (inferred) in doc 06.
  • No crash-injection test of the non-atomic save (ICX-AUD-047), and no dynamic Hermes forgery PoC (ICX-AUD-011). Both rest on reading the cited code.
  • Mutation testing covered icx-codec only (756 mutants). Other crates are a recommended next step.
  • Anthropic tokenization. Delimiter BPE alignment for Anthropic models was not measured.

Finding ID note

IDs run from ICX-AUD-001 to ICX-AUD-055. IDs 029 and 043–045 were reserved during drafting and merged into other findings (029 into 031; 043–045 into the "Positive results" sections). There are 51 findings in total.