safe-chains 0.222.0

Auto-allow safe bash commands in agentic coding tools
Documentation
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
# Filesystem region model (minimal, post-retreat).
#
# safe-chains classifies COMMANDS; it is deliberately NOT a filesystem trust database. Modeling
# "what every path on every machine means" is unbounded, per-machine, and drifts — and an
# adversarial audit of the old broad admit map (public /etc, /proc, /usr, /var/log, /Library)
# found it leaking (macOS keychain, Homebrew service configs, auth logs). Filesystem enforcement
# belongs to a sandbox / the OS / the harness, not a static string classifier (same scope stance
# as $PATH and hostile checkouts — see AGENTS.md §0.2).
#
# So this file is TINY and does only what a static classifier can honestly do:
#   • the WORKSPACE (relative paths under the harness cwd) is read/write — the agent was invited
#     there. That's the default `worktree` role; no node needed.
#   • a few `.git`/`.envrc`-style in-project files are read-safe but WRITE-frozen (a hook/direnv
#     write is code-injection).
#   • scratch (`/tmp`) and standard streams (`/dev/null`, fd pseudo-files) are read/write.
#   • a small CREDENTIAL SHIELD (`.ssh`, `.aws`, …) is denied and — crucially — a user grant can
#     NEVER widen it (see `apply_grant`), so `grant ~/` still can't hand over an SSH key.
#   • EVERYTHING ELSE (any absolute or `~` path matching nothing) → `unknown` → deny/abstain.
#     Reading `/etc/hosts` or `/usr/bin/python3` now prompts instead of auto-approving; a user
#     who wants those adds a read grant to ~/.config/safe-chains.toml. This is the retreat: we
#     stop guessing the filesystem and let the user's grants encode their own machine's truth.
#
# Matching (header rules unchanged): a `path` ending `/` = SUBTREE prefix; ending `*` = STRING
# prefix; no `/` = SEGMENT (matches any path component == it, at any depth); else EXACT. Most
# specific wins; fail-closed to `unknown`. Credential shields are SEGMENTS so they bite whatever
# the spelling (`~/.ssh/id_rsa`, a relative `.ssh/id_rsa`, a `.ssh/` inside a granted tree).
#
# PROVENANCE: every node keeps a `note` (why it earns its role) and a dated `researched`.

# ── Roles: role → { read face, write face, rebind face, secret, frozen } ───────────────────
#
# THREE faces, not two. `read_locus` and `write_locus` are the obvious ones. `rebind_locus` is the
# locus of an operation that changes what the NAME refers to rather than the bytes under it: `rm`
# and `rmdir` unbind it, `ln` points it elsewhere, `mv` takes it away. It defaults to the write
# face and only needs stating where the two must differ.
#
# `frozen` says which face a user GRANT may not widen:
#   (absent)          a grant widens every face
#   frozen = "rebind" a grant may write INTO the node but never replace it
#   frozen = "write"  a grant may not write it at all, and therefore cannot rebind it either
#
# The distinction is load-bearing rather than decorative. Freezing the WRITE face of `~/.config` to
# stop `ln -s /tmp/evil ~/.config` also denied `cp x ~/.config`, which is an ordinary thing to do in
# a directory you granted. Freezing only the REBIND face keeps both answers right. It cannot be
# derived from the operation, because `cp x DIR` and `ln -s y DIR` are both `create`/`transfer` to
# the engine — `ln` declares `rebinds_destination` in its own TOML to separate them.
[role.worktree]
description = "The project tree (relative paths under the workspace). Read + write."
read_locus = "worktree"
write_locus = "worktree"
reads_secret = false

[role.worktree-trusted]
description = "In-project files another tool auto-executes (.git internals, .envrc). Read-safe; writes frozen (a hook/config write is code-injection)."
read_locus = "worktree-trusted"
write_locus = "worktree-trusted"
reads_secret = false

[role.scratch]
description = "Process-scoped temporary scratch."
read_locus = "temp"
write_locus = "temp"
reads_secret = false

[role.standard-stream]
description = "Standard streams / fd pseudo-files — no real filesystem is touched."
read_locus = "process"
write_locus = "process"
reads_secret = false

[role.package-content]
description = "Distributed package CONTENT — docs, headers, libraries, vendored module sources. Read-admitted; writes stay denied."
read_locus = "adjacent"
write_locus = "machine"
reads_secret = false

[role.credential-store]
description = "Known credential/secret stores. Denied, and a user grant may NOT widen them (the shield)."
read_locus = "machine"
write_locus = "machine"
reads_secret = true

[role.safe-chains-config]
description = "safe-chains' own config. Read-OK, but write DENIED and un-grantable — an agent must not rewrite the file that governs what it may do. The write face sits at `system-integrity`, not `machine`: this file IS the classifier's trust substrate, and a level that admits ordinary machine administration must not thereby admit rewriting its own ceiling."
read_locus = "worktree-trusted"
write_locus = "system-integrity"
reads_secret = false
frozen = "write"

[role.trust-source]
description = "A file safe-chains derives its own trust from: a harness settings file it reads allow-rules or read-approvals out of. Write DENIED and un-grantable — an agent that rewrites one of these decides what safe-chains approves next. Unlike safe-chains' own config the READ face stays machine, because a harness settings file can carry an `env` block with API keys. Its write face sits at `system-integrity` for the same reason safe-chains own config does: a level admitting ordinary machine administration must not thereby admit rewriting the file that decides what gets approved next."
read_locus = "machine"
write_locus = "system-integrity"
reads_secret = false
frozen = "write"

[role.trust-root-dir]
description = "The DIRECTORY a trust file lives in. Writing a file into it is ordinary and stays grantable; REBINDING it is not, because replacing the directory points the trust root at somewhere else entirely and every file pin beneath it becomes irrelevant. A grant on such a directory therefore permits `cp x ~/.config` and refuses `rm -rf ~/.config` and `ln -s /tmp/evil ~/.config`. The REBIND face sits at `system-integrity` so the freeze survives a raised level, while the write face stays at `machine` and remains grantable."
read_locus = "machine"
write_locus = "machine"
rebind_locus = "system-integrity"
reads_secret = false
frozen = "rebind"

[role.unknown]
description = "Fail-closed default: any absolute/home path matching no region. Deny read & write (the harness then prompts)."
read_locus = "machine"
write_locus = "machine"
reads_secret = false

# ── In-project trust carve-outs: read-safe, write-frozen (code-injection) ───────────────────
[[region]]
path = ".git"
role = "worktree-trusted"
note = "Git metadata dir; hooks/ and config execute code on git operations, so a write is code-injection. Nests at any depth."
researched = "2026-07-12"
[[region]]
path = ".envrc"
role = "worktree-trusted"
note = "direnv auto-executes .envrc on cd into its directory; a write plants code that runs on the next shell entry."
researched = "2026-07-12"

# ── Scratch + standard streams: read/write ─────────────────────────────────────────────────
[[region]]
path = "/tmp/"
role = "scratch"
note = "FHS world-writable temp; ephemeral scratch. Both OSes (macOS /tmp symlinks to /private/tmp)."
researched = "2026-07-12"
[[region]]
path = "/var/tmp/"
role = "scratch"
note = "FHS temp preserved across reboots; still process-throwaway in intent."
researched = "2026-07-12"
[[region]]
path = "/private/tmp/"
role = "scratch"
os = ["macos"]
note = "macOS's real temp directory; /tmp is a symlink to here."
researched = "2026-07-12"
[[region]]
path = "/private/var/tmp/"
role = "scratch"
os = ["macos"]
note = "macOS real location of /var/tmp."
researched = "2026-07-12"
[[region]]
path = "/dev/null"
role = "standard-stream"
note = "POSIX bit-bucket; discards writes, empty on read."
researched = "2026-07-12"
[[region]]
path = "/dev/stdout"
role = "standard-stream"
note = "POSIX stdout pseudo-file — the process's own output stream."
researched = "2026-07-12"
[[region]]
path = "/dev/stderr"
role = "standard-stream"
note = "POSIX stderr pseudo-file — the process's own error stream."
researched = "2026-07-12"
[[region]]
path = "/dev/tty"
role = "standard-stream"
note = "The controlling terminal; process-local I/O."
researched = "2026-07-12"
[[region]]
path = "/dev/fd/"
role = "standard-stream"
note = "Per-process file-descriptor pseudo-files (process substitution). Process-local."
researched = "2026-07-12"

# ── Package CONTENT: read-admitted, writes still denied. ───────────────────────────────────────
#
# The retreat deleted a broad admit map because an adversarial audit found it leaking — macOS
# `/Library/Keychains/System.keychain`, Homebrew `/usr/local/etc/*` service configs, `/var/log/*`
# auth tokens. That map failed because it admitted whole ROOTS, and a root is HETEROGENEOUS:
# `/usr/local` holds distributed package files next to machine-local config and state.
#
# These nodes cut at the layer the FHS already separates. `share`/`include`/`lib` is content a
# PACKAGE MANAGER put there; `etc` and `var` are this machine's config and state, which is where
# every audited leak lived. None of the three audit findings matches a prefix below: `/Library`
# is admitted only at `Developer/CommandLineTools/`, `etc` is never admitted at any root, and
# `/var` is not a root here at all.
#
# The argument for reading these at all: the content is public by construction. A crate README
# under `~/.cargo/registry` is published on crates.io; a man page ships in the OS. Refusing the
# local copy while the same bytes are a network fetch away is friction without a matching risk.
# WRITES stay denied — this widens disclosure only, not the ability to alter what runs.
[[region]]
path = "/usr/share/"
role = "package-content"
note = "OS-distributed docs, man pages, locale data. Sibling `/usr/etc` does not exist; config lives in /etc, which is NOT admitted."
researched = "2026-07-30"
[[region]]
path = "/usr/include/"
role = "package-content"
note = "System C headers — compiler input, no machine state."
researched = "2026-07-30"
[[region]]
path = "/usr/lib/"
role = "package-content"
note = "OS-distributed libraries."
researched = "2026-07-30"
[[region]]
path = "/usr/local/share/"
role = "package-content"
note = "Locally-installed package docs/data. Deliberately NOT /usr/local/etc, which the admit-audit found holding Homebrew service secrets."
researched = "2026-07-30"
[[region]]
path = "/usr/local/include/"
role = "package-content"
note = "Locally-installed headers."
researched = "2026-07-30"
[[region]]
path = "/usr/local/lib/"
role = "package-content"
note = "Locally-installed libraries."
researched = "2026-07-30"
[[region]]
path = "/opt/homebrew/share/"
role = "package-content"
note = "Homebrew package docs/data. NOT /opt/homebrew/etc (service configs) and NOT Cellar/, whose per-formula trees contain their own etc/."
researched = "2026-07-30"
[[region]]
path = "/opt/homebrew/include/"
role = "package-content"
note = "Homebrew headers."
researched = "2026-07-30"
[[region]]
path = "/opt/homebrew/lib/"
role = "package-content"
note = "Homebrew libraries and node_modules for globally-installed packages."
researched = "2026-07-30"
[[region]]
path = "/Library/Developer/CommandLineTools/"
role = "package-content"
note = "Apple SDK headers and toolchain sources. Scoped to Developer/ so it cannot reach /Library/Keychains, which the admit-audit found leaking."
researched = "2026-07-30"
[[region]]
path = "/nix/store/"
role = "package-content"
note = "Nix store paths are immutable, content-addressed and world-readable by design."
researched = "2026-07-30"
[[region]]
path = "~/.cargo/registry/"
role = "package-content"
note = "Vendored crate sources fetched from a registry. Scoped to registry/ so it cannot reach ~/.cargo/credentials.toml."
researched = "2026-07-30"
[[region]]
path = "~/.rustup/toolchains/"
role = "package-content"
note = "Installed Rust toolchains and their std sources."
researched = "2026-07-30"
[[region]]
path = "~/go/pkg/mod/"
role = "package-content"
note = "Go module cache — downloaded dependency sources."
researched = "2026-07-30"
[[region]]
path = "~/.nvm/versions/"
role = "package-content"
note = "nvm-installed node runtimes and their global node_modules."
researched = "2026-07-30"
[[region]]
path = "~/.local/bin/"
role = "package-content"
note = "Installed user-level executables (mise/pipx/cargo-install shims). Reading one — `strings`, `cat` on a wrapper script — discloses an installed tool, not machine state. Writes stay denied: this is where a planted binary would go."
researched = "2026-07-30"
[[region]]
path = "~/.local/share/mise/installs/"
role = "package-content"
note = "mise-installed tool versions. Scoped to installs/ so it cannot reach mise's own config or state."
researched = "2026-07-30"

# ── Credential shield: SEGMENT matches, denied everywhere, and NOT widenable by a grant. The
#    only "denylist" we keep, and only because a broad grant (`grant ~/`) would otherwise hand
#    over these. A project that literally contains one of these dirs is denied too — rare and
#    acceptable. Everything else in home is already deny-by-default; these just can't be granted. ─
[[region]]
path = ".ssh"
role = "credential-store"
note = "SSH private keys / authorized_keys / known_hosts — credential + trust store."
researched = "2026-07-12"
[[region]]
path = ".aws"
role = "credential-store"
note = "AWS access keys / session credentials."
researched = "2026-07-12"
[[region]]
path = ".gnupg"
role = "credential-store"
note = "GnuPG private keyring and trust database."
researched = "2026-07-12"
[[region]]
path = ".kube"
role = "credential-store"
note = "Kubernetes kubeconfig — cluster credentials and tokens."
researched = "2026-07-12"
[[region]]
path = ".docker"
role = "credential-store"
note = "Docker registry auth tokens / credential-helper config."
researched = "2026-07-12"
[[region]]
path = ".netrc"
role = "credential-store"
note = "Plaintext machine login credentials for ftp/curl/git."
researched = "2026-07-12"

# ── System credential stores + other-user homes: absolute paths a broad grant (`grant /` or
#    `grant /etc/`) would otherwise expose. Kept as un-grantable shields (the dotfile rule only
#    covers hidden entries; these are not dotfiles). ──────────────────────────────────────────
[[region]]
path = "/etc/shadow"
role = "credential-store"
os = ["linux"]
note = "Hashed user passwords — the credential store proper."
researched = "2026-07-12"
[[region]]
path = "/etc/gshadow"
role = "credential-store"
os = ["linux"]
note = "Hashed group passwords."
researched = "2026-07-12"
[[region]]
path = "/etc/master.passwd"
role = "credential-store"
os = ["macos"]
note = "macOS password hash store."
researched = "2026-07-12"
[[region]]
path = "/etc/ssl/private/"
role = "credential-store"
note = "TLS/service private keys."
researched = "2026-07-12"
[[region]]
path = "/root/"
role = "credential-store"
os = ["linux"]
note = "root's home — privileged private data + credentials."
researched = "2026-07-12"
[[region]]
path = "/var/root/"
role = "credential-store"
os = ["macos"]
note = "macOS root home."
researched = "2026-07-12"
# Well-known credential stores NESTED under ~/.config (non-hidden, so the dotfile rule can't
# catch them under an explicit `grant ~/.config/`). Not exhaustive — an explicit grant of a
# config dir can still expose an unlisted app's secrets; the docs advise granting narrowly.
[[region]]
path = "~/.config/gh/"
role = "credential-store"
note = "GitHub CLI OAuth token (hosts.yml)."
researched = "2026-07-12"
[[region]]
path = "~/.config/gcloud/"
role = "credential-store"
note = "Google Cloud SDK credentials / tokens."
researched = "2026-07-12"
[[region]]
path = "~/.config/rclone/"
role = "credential-store"
note = "rclone remote credentials (rclone.conf)."
researched = "2026-07-12"
[[region]]
path = "~/.config/git/"
role = "credential-store"
note = "git's `store` credential helper — plaintext https://user:token@host."
researched = "2026-07-12"

# ── macOS ~/Library credential stores. The dotfile rule keeps a broad `grant ~/` from widening
#    HIDDEN entries, but on macOS the keychain, browser password stores, cookies, and the iMessage
#    DB live under ~/Library — NOT dot-prefixed — so a broad grant would otherwise sweep them up.
#    Shielded (un-grantable) like the credential dotdirs. Not exhaustive (Application Support holds
#    countless app secrets); the docs advise granting narrowly. ─────────────────────────────────
[[region]]
path = "~/Library/Keychains/"
role = "credential-store"
os = ["macos"]
note = "macOS keychain databases (login.keychain-db)."
researched = "2026-07-12"
[[region]]
path = "~/Library/Cookies/"
role = "credential-store"
os = ["macos"]
note = "Per-app HTTP cookie stores (session tokens)."
researched = "2026-07-12"
[[region]]
path = "~/Library/Safari/"
role = "credential-store"
os = ["macos"]
note = "Safari history / autofill / per-site data."
researched = "2026-07-12"
[[region]]
path = "~/Library/Messages/"
role = "credential-store"
os = ["macos"]
note = "iMessage database (chat.db) — private message history."
researched = "2026-07-12"
[[region]]
path = "~/Library/Application Support/Firefox/"
role = "credential-store"
os = ["macos"]
note = "Firefox profiles: logins.json + key4.db (offline-decryptable saved passwords)."
researched = "2026-07-12"
[[region]]
path = "~/Library/Application Support/Google/Chrome/"
role = "credential-store"
os = ["macos"]
note = "Chrome profile: Login Data / Cookies stores."
researched = "2026-07-12"
[[region]]
path = "~/Library/Application Support/Chromium/"
role = "credential-store"
os = ["macos"]
note = "Chromium profile credential/cookie stores."
researched = "2026-07-12"
[[region]]
path = "~/Library/Application Support/BraveSoftware/"
role = "credential-store"
os = ["macos"]
note = "Brave profile credential/cookie stores."
researched = "2026-07-12"
[[region]]
path = "~/Library/Application Support/Microsoft Edge/"
role = "credential-store"
os = ["macos"]
note = "Edge profile credential/cookie stores."
researched = "2026-07-12"

# ── safe-chains' own config: read-OK, write-DENIED, un-grantable (the trust root). An
#    UNPROTECTED path (no harness blesses a third-party config location) — a stopgap until one
#    exists; a write here is best-effort caught (a python/editor write escapes the command
#    classifier). This is the ONLY location safe-chains reads its config from — `XDG_CONFIG_HOME`
#    is not honored (an agent-mutable env var must not relocate the trust root). See the
#    protected-config research item. ────────────────────────────────────────────────────────────
[[region]]
path = "~/.config/safe-chains.toml"
role = "safe-chains-config"
note = "safe-chains' user config (grants). Writing it would let an agent grant itself permissions."
researched = "2026-07-12"

[[region]]
path = "~/.claude/settings.json"
role = "trust-source"
note = "safe-chains reads `permissions.allow` command patterns AND Read() path approvals out of this file (src/allowlist.rs, regions.rs::claude_settings_read_grants), and the hook gates on the coverage those patterns produce. Writing it is therefore a way to decide what safe-chains approves next, exactly like writing safe-chains' own config."
researched = "2026-08-02"

# ── Trust-root DIRECTORIES. Freezing the trust FILES is not enough on its own: a grant on the
#    parent let an agent delete the directory and replace it with a symlink to one it controls,
#    after which safe-chains read its grants (and its `level` ceiling) from the agent's own file.
#    Every step of that was auto-approved. safe-chains classifies a path by its literal spelling and
#    does not follow symlinks (AGENTS.md §0.2), which is right for classification but means the
#    RELOCATION has to be stopped where it happens, at the point the directory is replaced.
#
#    EXACT matchers, so only the directory node itself is frozen. Files underneath are untouched and
#    a `~/.config` grant still works normally for every other tool's config living there.
#    `/etc` needs no equivalent: it already denies as `machine`.
[[region]]
path = "~/.config"
role = "trust-root-dir"
note = "Parent of safe-chains' own config. `rm -rf ~/.config` then `ln -s /tmp/evil ~/.config` relocated the whole trust root."
researched = "2026-08-02"

[[region]]
path = "~/.claude"
role = "trust-root-dir"
note = "Parent of the harness settings file safe-chains reads allow-rules from. Same relocation."
researched = "2026-08-02"

# ── System-integrity substrate: the machine's own identity/auth/boot/loader files, where a WRITE is
#    compromise-complete — add a UID-0 user, grant yourself sudo, rewrite the auth/login path, redirect
#    the dynamic loader, or change the boot path. The READ face is ordinary machine config; the WRITE
#    face worst-cases to `system-integrity`, which sits ABOVE `machine` (and above local-admin), so
#    owning the trust substrate is a yolo-level grant — distinct from ordinary machine admin (a service
#    restart, an app-config edit). Deny-ward and fail-closed like the credential shield above: this is
#    NOT an admit map (the auto-approve band already denies all of `machine`); it only splits the
#    upper-level bands so "run the machine as admin" (local-admin) and "own the machine's trust root"
#    (yolo) are different grants. The secret-READ stores (/etc/shadow, master.passwd) stay in the
#    credential shield; these are the write-compromise files that are not themselves secrets. ────────
[role.system-integrity]
description = "The machine's identity/auth/boot/loader substrate. Readable as ordinary config; a WRITE is compromise-complete and worst-cases above local-admin (system-integrity)."
read_locus = "machine"
write_locus = "system-integrity"
reads_secret = false

[[region]]
path = "/etc/passwd"
role = "system-integrity"
note = "System account database; a write can add a UID-0 user — full compromise."
researched = "2026-07-16"
[[region]]
path = "/etc/group"
role = "system-integrity"
note = "System group database; a write can grant membership in privileged groups (wheel/sudo/admin)."
researched = "2026-07-16"
[[region]]
path = "/etc/sudoers"
role = "system-integrity"
note = "sudo policy; a write grants arbitrary root — the ultimate escalation."
researched = "2026-07-16"
[[region]]
path = "/etc/sudoers.d/"
role = "system-integrity"
note = "Drop-in sudo policy fragments; same grant-root power as /etc/sudoers."
researched = "2026-07-16"
[[region]]
path = "/etc/pam.d/"
role = "system-integrity"
note = "PAM authentication stack (Linux and macOS); a write rewrites how every login authenticates."
researched = "2026-07-16"
[[region]]
path = "/etc/ld.so.conf"
role = "system-integrity"
os = ["linux"]
note = "Dynamic-loader search config; a write can redirect every program to attacker libraries."
researched = "2026-07-16"
[[region]]
path = "/etc/ld.so.conf.d/"
role = "system-integrity"
os = ["linux"]
note = "Dynamic-loader search drop-ins; same library-hijack power as ld.so.conf."
researched = "2026-07-16"
[[region]]
path = "/etc/ld.so.preload"
role = "system-integrity"
os = ["linux"]
note = "Force-loaded shared objects for every dynamically-linked program — a machine-wide code-injection hook."
researched = "2026-07-16"
[[region]]
path = "/boot/"
role = "system-integrity"
os = ["linux"]
note = "Kernel images, initramfs, and bootloader config; a write changes what the machine boots."
researched = "2026-07-16"