Leia em Português

• The latest release receives security fixes, and older releases receive best-effort patches for critical vulnerabilities only
• Upgrade to the latest version before reporting, because a fix lands on the current line first
• The table lists the release line, and crates.io carries 0.1.37 while 0.1.38 is the version in preparation, so the pin in this document says which one is under test
• The versions 0.1.15, 0.1.16, 0.1.17, 0.1.21 and 0.1.33 were never published, so they are not listed

• Report security vulnerabilities privately by email to daniloaguiarbr@proton.me, which is the channel that works today
• GitHub Security Advisories through the repository Security tab is the second channel, available once the repository is published
• Do NOT open a public issue for security vulnerabilities
• Include: atomwrite version, OS, Rust version, description of the vulnerability, steps to reproduce, potential impact, proof of concept if available

• These are best-effort targets of a single maintainer, not a contractual SLA: acknowledgment within 48 hours, severity classification within 5 business days and a status update every 7 days
• A report that misses a target is still handled, and the reporter is told which target was missed

• These are best-effort targets, and severity follows CVSS 3.1: Critical 9.0-10.0, High 7.0-8.9, Medium 4.0-6.9, Low below 4.0
• Critical severity: patch within 7 days
• High severity: patch within 14 days
• Medium severity: patch within 30 days
• Low severity: patch in the next scheduled release

• Coordinated disclosure: fix first, disclose after
• Reporter is credited unless they request anonymity
• Public disclosure occurs after the fix is published to crates.io
• Disclosure includes: CVE (if applicable), affected versions, fixed version, description, mitigation

• Security patches are released as point releases (e.g., 0.1.13, 0.1.14)
• Announcements are posted via GitHub Security Advisories
• Users should subscribe to repository notifications for timely updates

• Status: resolved in v0.1.7 (2026-06-05)
• Issue: DoS via stack exhaustion in time parsing
• Fix: upgraded time to 0.3.47 with DEPTH_LIMIT=32, and the fix requires Rust 1.88
• Action taken: MSRV bumped from 1.85 to 1.88 in v0.1.7, the ignore entry in deny.toml and the cargo audit --ignore flag were both removed, and the advisory no longer applies
• Reference: CHANGELOG.md v0.1.7 entry

• Severity: Critical
• Affected versions: v0.1.0 through v0.1.26
• Fixed in: v0.1.27
• Attack vector: create a symlink to a directory outside the workspace (ln -s /tmp $WS/link), then use atomwrite to write/read through the symlink (atomwrite write link/escape.txt)
• Root cause: validate_path in path_safety.rs used soft_canonicalize which resolved . and .. but NOT symlinks in intermediate path components
• Impact: arbitrary file write and read outside the workspace jail (data exfiltration, arbitrary file creation)
• Affected subcommands: write, read, edit, edit-loop, apply, set, del, copy, move
• NOT affected: search, replace (use WalkBuilder from the ignore crate with own symlink verification)
• Fix: added canonicalize_existing_prefix that resolves symlinks via std::fs::canonicalize on the existing portion of the path before jail verification
• Mitigation: upgrade to v0.1.27 and do not create symlinks inside the workspace pointing outside

• The audit at v0.1.36 was not re-run before that release, and running it on the v0.1.37 tree found 7 vulnerabilities, 2 of them high (7.5) in quick-xml 0.37
• Both high ones were reachable from extract-artifact reading XML from an OOXML container: unbounded namespace allocation (RUSTSEC-2026-0195) and quadratic duplicate-attribute checking (RUSTSEC-2026-0194), fixed by moving to quick-xml 0.42
• zip 9.0.5 closes RUSTSEC-2026-0367, RUSTSEC-2026-0368 and RUSTSEC-2026-0369 (one of them critical, 9.1) in the OOXML container reader
• crossbeam-epoch, rustls, anyhow and memmap2 were updated to their patched versions in the same pass
• Lesson recorded: run cargo audit and cargo deny check BEFORE the version bump, not after the docs are written

• cargo audit on Cargo.lock (566 crate dependencies, 1403 advisories in the local database, last updated 2026-11-03): exit 0, no vulnerability and no warning, measured on 2026-11-04
• cargo deny check: advisories ok, bans FAILED, licenses FAILED, sources ok, exit 6, measured on 2026-11-04 on this tree, and the failures are named in the section below
• The unmaintained-crate notices of the earlier audits, instant (RUSTSEC-2024-0384) and unsafe-libyaml (RUSTSEC-2024-0449), are gone: neither those two nor serde_yaml or serde_yaml_ng is in Cargo.lock, and cargo tree --locked --all-features -i <crate> answers "did not match any packages" for each
• The YAML rules of transform are read with serde-saphyr 1.3.0 and notify is on 8.x (ADR-0057 sections 4 and 5)
• Re-run both tools on the final tree before every release, the numbers above are a measurement of one day

• cargo deny check exited 6 on 2026-11-04, and the failing checks are bans and licenses
• heic 0.1.6 is a direct dependency of this crate and declares AGPL-3.0-only OR LicenseRef-Imazen-Commercial, so both halves of the expression are rejected by the deny.toml allowlist
• AGPL-3.0 is a copyleft licence, so publishing this tree to crates.io would put an AGPL dependency under every consumer of the binary
• The remaining two failures are duplicate entries the 33 skip rows do not carry: fearless_simd 0.7.0 next to 1.2.0 (through hayro -> vello_cpu) and the duplicate pairs of syn and weezl
• The gate stays red until heic is removed, replaced by a permissively licensed decoder, or the owner records a written exception with the reason, which the allowlist rule requires to be explicit

• read --ocr fetches two models from huggingface.co/robertknight/ocrs with a 64 MiB cap and a 120 second timeout, and neither a checksum nor a signature is verified before the file is renamed into the cache
• The tree-sitter bundle checks the sha256 listed in its manifest; the OCR path has no equivalent, so a compromised host or a proxy in the middle can serve a different model and the product loads it in-process
• Mitigation today: pre-populate $XDG_CACHE_HOME/atomwrite/models/ from a trusted machine, and keep that directory writable by your user only
• The fix is to pin one sha256 per model in src/constants.rs and refuse a file that does not match, which is open work and is not part of 0.1.38

• --old-file / --new-file content-source paths remain under the workspace jail (no OS /tmp allowlist)
• Agents MUST stage patch bodies via atomwrite scratch init .atomwrite/scratch under $workspace/.atomwrite/scratch/ then pass those jail-safe paths
• A WORKSPACE_JAIL (exit 126) on /tmp/... old/new files is intended policy (v0.1.36 / ADR-0055), not a security hole to bypass

• Memory safety: unsafe is denied crate-wide (#![deny(unsafe_code)] in src/lib.rs and src/main.rs) and allowed only in src/signal.rs and src/platform.rs (#![allow(unsafe_code)] at line 5 of each), where 8 unsafe blocks wrap signal-hook registration and libc calls (_exit, poll, write to stderr, fcntl on macOS, renameat2 on Linux, console setup on Windows)
• The 8 blocks are at src/signal.rs:260, src/signal_handlers.inc.rs:31,143,171,219 (included from src/signal.rs) and src/platform.rs:97,203,275, counted with rg 'unsafe\s*\{' src, which returns 9 matches because src/commands/scope/queries.inc.rs:77 holds the literal text of a query pattern, and each block carries a SAFETY: comment
• BLAKE3: used for checksums only, not for cryptographic security
• tree-sitter-language-pack 1.22.1: parsers are downloaded on first use from the release assets of github.com/xberg-io/tree-sitter-language-pack through the download feature (HTTP client ureq with rustls and ring, all in the default binary)
• The client checks every bundle against the sha256 listed in the manifest fetched from the same release (download.rs of the crate), and a bundle that fails the check is rejected
• The downloaded parsers are native libraries loaded into the process (dynamic-loading), so they run as code: keep the cache directory ($XDG_CACHE_HOME/tree-sitter-language-pack/) writable only by your user
• To run with no download, pre-populate that cache or build without ast, see PRIVACY.md
• deny.toml: includes MPL-2.0, CDLA-Permissive-2.0, CC0-1.0 in the allowlist, and has 33 skip entries, each with a reason, for the unavoidable coexistence of duplicate crate versions (getrandom 0.2/0.3, rustix 0.x/1.x, windows-sys 0.52/0.59 and the Windows crates of the opt-in watch feature)
• MSRV: Rust 1.90 stable (rust-version in Cargo.toml, raised from 1.89 in v0.1.38, see ADR-0075)

• Security researchers who report valid vulnerabilities are recognized here
• Report to be listed (or request anonymity)
• To request listing, open a GitHub Security Advisory with the report and include your preferred attribution

• Use --workspace to restrict operations to the project root
• Avoid running atomwrite as root
• Validate --expect-checksum in multi-agent environments
• Review NDJSON error output for ok: false, retryable, and suggestion fields (exit-first parse)
• A bounded read answers truncated: true and a next range, so never treat a truncated read as the whole file, because content past the cap stays invisible to the review
• On network filesystems flock can be silently ignored, so keep --expect-checksum as the extra net there, and remember the sidecar is best effort while the inode lock carries a read-only directory
• Keep atomwrite updated to the latest version
• Audit the batch manifest before execution in production
• Subscribe to repository notifications for timely security updates
• Treat orphan journals (.atomwrite.journal.* sidecars from a previous crash) with suspicion: inspect the target file AND the journal content before deleting the journal
• When --syntax-check is enabled, do NOT pipe sensitive content through stdin on shared systems: the error envelope may echo source location and surrounding context
• Stage agent patches with scratch init + --old-file/--new-file under the workspace, and do not expect OS /tmp to be jail-exempt
• A crash report (report-<uuid>.toml in the OS temporary directory, written by the human-panic hook and never transmitted) can hold paths and a backtrace, so review it before you attach it to an issue