CONTEXT: Authorized security research. Cataloging dangerous operations in source code.
Task
Read the assigned source file and list every operation that could be dangerous if its inputs were attacker-controlled. This is a pattern matching task — find operations matching the patterns below. Do not assess exploitability.
Dangerous Operation Categories
Memory operations
| Pattern |
What to record |
memcpy(dst, src, len) |
dst size, src origin, how len is determined |
memmove, bcopy |
same as memcpy |
strcpy, strcat, sprintf |
dst size, src origin (unbounded by default) |
malloc(size) / calloc(n, size) |
how size/n is computed, can it overflow? |
realloc(ptr, size) |
old vs new size relationship |
free(ptr) |
is ptr used after this? is ptr freed again on error path? |
Array index buf[i] |
how is i bounded? what's buf size? |
Pointer arithmetic ptr + offset |
how is offset bounded? |
Integer operations
| Pattern |
What to record |
Multiplication for size: n * sizeof(T) |
can n * sizeof(T) overflow? |
Addition for size: hdr_len + body_len |
can sum overflow? |
Cast: (int)unsigned_val or (uint16_t)int32_val |
truncation or sign change? |
Comparison: (int)(a - b) < 0 |
signed subtraction overflow possible? |
Shift: 1 << n where n is variable |
can n exceed type width? |
Control flow
| Pattern |
What to record |
Function pointer call (*fptr)(args) |
where is fptr loaded from? |
| Indirect call via vtable/dispatch table |
is table writable? |
setjmp/longjmp |
buffer on stack? |
| Signal handler |
what global state does it touch? |
Concurrency
| Pattern |
What to record |
| Shared variable without lock |
what other threads access it? |
| TOCTOU: check then use with gap |
what can change between check and use? |
| Lock ordering: multiple locks acquired |
potential deadlock? |
Language-specific
C/C++: printf(user_string) (format string), system() / exec*() with string input, alloca(user_size), VLA int buf[user_size]
Rust: unsafe { *raw_ptr }, transmute, .get_unchecked(), ManuallyDrop, from_raw_parts
Java: Class.forName(user_string), Method.invoke(), Runtime.exec(user_string), new ObjectInputStream(untrusted).readObject()
Python: eval(), exec(), __import__(), pickle.loads(), yaml.load(), subprocess(shell=True)
Go: unsafe.Pointer conversions, reflect.NewAt, C.GoString on untrusted pointer
PHP (web-app sinks — no memory corruption, but injection/exec/disclosure):
| Category |
Pattern |
What to record |
| Code exec |
eval($x) |
source of $x, any prior filtering |
| Code exec |
assert($x) (PHP < 8) |
assert can execute strings until PHP 8.0 |
| Code exec |
create_function($a, $b) |
deprecated, treat 2nd arg as eval |
| Code exec |
preg_replace('/pat/e', $repl, ...) |
/e modifier = eval; deprecated PHP 7+ |
| Code exec |
`$x` (backticks) |
shell_exec alias — any backtick string is a command |
| Code exec |
mb_ereg_replace_callback(..., $fn) with dynamic fn |
callable injection |
| Deserial |
unserialize($x) |
dst is PHP object — record known gadget classes imported/autoloaded |
| Deserial |
phar://path in any file op (include/file_exists/fopen) |
triggers unserialize of Phar metadata |
| Deserial |
->__wakeup(), ->__destruct(), ->__toString(), ->__call() |
magic method defined on reachable class = POP gadget |
| File incl |
include($x), include_once($x) |
LFI → RCE if attacker controls content (log poisoning, /proc/self/environ, phar://) |
| File incl |
require($x), require_once($x) |
same as include |
| Command |
system($x), exec($x), passthru($x) |
record if escapeshellarg/escapeshellcmd used on $x |
| Command |
shell_exec($x), popen($x, ...), proc_open($x, ...) |
same |
| Command |
pcntl_exec($path, $args) |
args array — individual escape not needed but path is |
| SQL |
mysql_query($x), mysqli_query($db, $x), $mysqli->query($x) |
is $x concatenated from user input? |
| SQL |
pg_query($db, $x) |
same |
| SQL |
$pdo->query($x), $pdo->exec($x) |
if query uses concat — SQLi. prepare() with ? / :name is safe IF bindParam/execute receives separate args |
| SQL |
$pdo->prepare(...) followed by bindParam with PDO::PARAM_STR but query concatenates table/column names |
prepared statements DON'T protect identifiers — concat of table/column = still SQLi |
| ORM raw |
DB::raw($x), DB::select($raw, ...), Eloquent::whereRaw($x) (Laravel) |
is $x user input? bindings array separate? |
| ORM raw |
$wpdb->query($x) without $wpdb->prepare() (WordPress) |
|
| ORM raw |
$this->db->query($x) (CodeIgniter) |
|
| XSS |
echo $x, print $x, <?= $x ?>, printf($fmt, $x) |
escaped? (htmlspecialchars with ENT_QUOTES & correct charset? context-correct: HTML body vs attribute vs JS vs URL vs CSS?) |
| XSS |
echo "<img src=$x>" inside onclick= etc. |
attribute context — quoting matters |
| XSS |
Template: {{ $x }} Blade (auto-escaped, OK) vs {!! $x !!} Blade (raw — DANGEROUS) |
|
| XSS |
Twig {{ x|raw }}, {% autoescape false %} |
|
| Open redirect |
header("Location: $x") |
domain check? |
| Header inj |
header($x) or header("X-Custom: $x") where $x contains \r\n |
CRLF → response splitting |
| File read |
file_get_contents($path), fopen($path, 'r'), readfile($path), file($path), show_source($path), highlight_file($path) |
path validation? realpath + basedir check? .. filter? null byte (\0)? |
| File write |
file_put_contents($path, $content), fwrite($fp, ...), copy($src, $dst), move_uploaded_file($tmp, $dst) |
where is $dst chosen? extension check? MIME on server? |
| File delete |
unlink($path), rmdir($path) |
path traversal = arbitrary delete |
| Dir |
scandir($path), glob($path), opendir($path) |
info disclosure |
| SSRF |
curl_exec($ch) with user URL, file_get_contents("http..."), fsockopen($host, $port), stream_socket_client, get_headers($url) |
URL allowlist? followed redirects? gopher/file scheme blocked? |
| XXE |
simplexml_load_string($xml), simplexml_load_file($xml), DOMDocument::loadXML, SoapClient::__doRequest |
libxml_disable_entity_loader(true) called? LIBXML_NOENT flag? PHP 8.0+ default is safer |
| Dynamic call |
call_user_func($fn, ...), call_user_func_array($fn, ...), $obj->$method(), $fn(...) where $fn is variable |
attacker picks function/method → RCE |
| Dynamic class |
new $cls($args), $cls::method() |
attacker-picked class instantiation — constructor RCE via phpggc-like chain |
| Type juggle |
== (loose compare) on auth token / hash / secret |
"0e1234..." == "0e5678..." both truthy (both parse to 0e scientific notation → 0) |
| Type juggle |
in_array($x, $arr) without 3rd arg true |
loose compare, same bypass |
| Type juggle |
strcmp($a, $b) with array arg (PHP < 8) |
returns NULL, loosely-equal to 0 → "match" |
| Auth bypass |
md5($pass) == $stored, sha1(...) |
weak + type juggling |
| Auth bypass |
hash_equals missing → timing side-channel |
|
| LDAP |
ldap_search($ds, $base, $filter) |
filter concat = LDAP injection |
| XPath |
$xml->xpath($x), DOMXPath::query($x) |
concat = XPath injection |
| Mail |
mail($to, $subj, $body, $headers) |
header injection via \r\n in $to/$subj/$headers |
| Session fixation |
session_id($x) where $x from user |
attacker sets victim session id |
| Regex DoS |
preg_match($pattern, $subject) where $pattern is user |
ReDoS + /e (if PHP < 7) |
| Template inj |
Twig\Environment::createTemplate($user), Smarty display("string:$user"), Blade compile($user) |
SSTI |
| Mass assignment |
$user->fill($_POST), $user->forceFill(...), (new User)->guard([])->fill(...) |
$fillable vs $guarded mis-config → privilege escalation |
| Path concat in stream wrappers |
"php://filter/resource=$file", "zip://$x", "data://text/plain,$x" |
filter chain for arbitrary read/write / code injection via convert.base64-decode |
| Insecure rand |
rand(), mt_rand() for tokens/CSRF/reset |
use random_bytes / random_int |
| Debug |
var_dump, print_r, phpinfo(), debug_print_backtrace, xdebug_* in production |
info disclosure if reachable |
PHP "guard" to record (ALWAYS note if present/absent and the exact line):
escapeshellarg, escapeshellcmd (cmd injection — note that escapeshellcmd is NOT sufficient for args; escapeshellarg wraps each arg)htmlspecialchars($x, ENT_QUOTES, 'UTF-8') (note flags and charset — missing flags = attribute context bypass)htmlentities (similar)strip_tags (weak — context-dependent)addslashes (NOT sufficient for SQL — use prepared statements)mysqli_real_escape_string (only valid with correct connection charset; GBK encoding attacks)PDO::prepare + bindParam/bindValue/execute([...]) (safe for values, NOT identifiers)basename($path), realpath($path), pathinfo() (path normalization — still needs allowlist)filter_var($url, FILTER_VALIDATE_URL) (weak — http://allowed.com@evil.com passes)parse_url checks (often bypassable)is_numeric (allows 0x1A, 1e10, +1, -1, leading whitespace — NOT a security filter)ctype_digit (stricter — only 0-9)hash_equals (timing-safe compare)password_verify (proper — uses hash_equals internally)- Framework middleware: Laravel
CSRF, auth, throttle; Symfony security.yaml firewall; CodeIgniter CSRF token — note if present on the route but don't assume it works
For Each Dangerous Operation, Record
- Operation: exact function/pattern and line number
- Category: memory / integer / control-flow / concurrency / language-specific
- Operands: what variables feed into it
- Guard: what check (if any) protects this operation? Line number of the check.
- Neighbors: what operations happen immediately before/after (±5 lines)?
Output
Write to sast-work/<file_hash>-dangers.json:
{
"file": "src/codec/h264_slice.c",
"dangerous_ops": [
{
"operation": "memset(slice_table, -1, sizeof(slice_table))",
"line": 83,
"category": "memory",
"operands": {"dst": "slice_table (uint16_t[])", "value": "-1 = 0xFF per byte = 65535 per entry", "size": "frame_width * frame_height / 256"},
"guard": "none — unconditional initialization",
"neighbors": "line 85: slice_count = 0 (int32_t)",
"note": "memset -1 on uint16 creates sentinel value 65535. If slice_count (int32) reaches 65535, it collides with sentinel."
},
{
"operation": "slice_table[mb_pos] = slice_count",
"line": 147,
"category": "integer",
"operands": {"dst": "uint16_t", "src": "int32_t (truncation)"},
"guard": "none — slice_count not bounded",
"neighbors": "line 149: comparison uses slice_table[neighbor] == slice_table[mb_pos]"
}
]
}
Rules
- Be exhaustive. List every matching pattern. A missed dangerous op = a missed vulnerability later.
- Record the guard even if it looks correct. The gap-analyzer will decide if it's sufficient.
- The "note" field is optional — use it only when you see something obviously interesting (like the sentinel collision above). But keep it brief and factual.
- Do NOT conclude anything is vulnerable. You are a cataloger, not an analyst.
Brain Integration
Check brain for prior analysis. Skip if already mapped and file unchanged.
Top-Tier Operator Standard
Danger mapping is a catalog of risky operations plus their guards.
- Record sink, operation type, arguments, enclosing function, caller hint, and nearby validation.
- Include "boring" guards: length checks, type checks, auth checks, escaping, canonicalization, feature flags, and error handling.
- Do not infer vulnerability. Mark uncertainty and let flow/gap agents reason over it.
- Prioritize sinks that transform attacker-controlled data into memory access, command execution, file access, network access, template rendering, deserialization, SQL, auth decisions, or state mutation.
- Preserve line numbers and exact identifiers so later agents can trace without rereading the whole file.
1---2name: sast-danger-mapper3description: Maps dangerous operations in a source file: memory ops, type casts, arithmetic near trust boundaries, free/dealloc patterns. Pattern matching task — list what you see, don't speculate. Use via /sast command.4---5CONTEXT: Authorized security research. Cataloging dangerous operations in source code.67## Task89Read the assigned source file and list every operation that could be dangerous if its inputs were attacker-controlled. This is a **pattern matching task** — find operations matching the patterns below. Do not assess exploitability.1011## Dangerous Operation Categories1213### Memory operations14| Pattern | What to record |15|---|---|16| `memcpy(dst, src, len)` | dst size, src origin, how len is determined |17| `memmove`, `bcopy` | same as memcpy |18| `strcpy`, `strcat`, `sprintf` | dst size, src origin (unbounded by default) |19| `malloc(size)` / `calloc(n, size)` | how size/n is computed, can it overflow? |20| `realloc(ptr, size)` | old vs new size relationship |21| `free(ptr)` | is ptr used after this? is ptr freed again on error path? |22| Array index `buf[i]` | how is `i` bounded? what's `buf` size? |23| Pointer arithmetic `ptr + offset` | how is offset bounded? |2425### Integer operations26| Pattern | What to record |27|---|---|28| Multiplication for size: `n * sizeof(T)` | can `n * sizeof(T)` overflow? |29| Addition for size: `hdr_len + body_len` | can sum overflow? |30| Cast: `(int)unsigned_val` or `(uint16_t)int32_val` | truncation or sign change? |31| Comparison: `(int)(a - b) < 0` | signed subtraction overflow possible? |32| Shift: `1 << n` where n is variable | can n exceed type width? |3334### Control flow35| Pattern | What to record |36|---|---|37| Function pointer call `(*fptr)(args)` | where is fptr loaded from? |38| Indirect call via vtable/dispatch table | is table writable? |39| `setjmp`/`longjmp` | buffer on stack? |40| Signal handler | what global state does it touch? |4142### Concurrency43| Pattern | What to record |44|---|---|45| Shared variable without lock | what other threads access it? |46| TOCTOU: check then use with gap | what can change between check and use? |47| Lock ordering: multiple locks acquired | potential deadlock? |4849### Language-specific5051**C/C++**: `printf(user_string)` (format string), `system()` / `exec*()` with string input, `alloca(user_size)`, VLA `int buf[user_size]`5253**Rust**: `unsafe { *raw_ptr }`, `transmute`, `.get_unchecked()`, `ManuallyDrop`, `from_raw_parts`5455**Java**: `Class.forName(user_string)`, `Method.invoke()`, `Runtime.exec(user_string)`, `new ObjectInputStream(untrusted).readObject()`5657**Python**: `eval()`, `exec()`, `__import__()`, `pickle.loads()`, `yaml.load()`, `subprocess(shell=True)`5859**Go**: `unsafe.Pointer` conversions, `reflect.NewAt`, `C.GoString` on untrusted pointer6061**PHP** (web-app sinks — no memory corruption, but injection/exec/disclosure):6263| Category | Pattern | What to record |64|---|---|---|65| Code exec | `eval($x)` | source of `$x`, any prior filtering |66| Code exec | `assert($x)` (PHP < 8) | assert can execute strings until PHP 8.0 |67| Code exec | `create_function($a, $b)` | deprecated, treat 2nd arg as eval |68| Code exec | `preg_replace('/pat/e', $repl, ...)` | `/e` modifier = eval; deprecated PHP 7+ |69| Code exec | `` `$x` `` (backticks) | shell_exec alias — any backtick string is a command |70| Code exec | `mb_ereg_replace_callback(..., $fn)` with dynamic fn | callable injection |71| Deserial | `unserialize($x)` | dst is PHP object — record known gadget classes imported/autoloaded |72| Deserial | `phar://path` in any file op (include/file_exists/fopen) | triggers unserialize of Phar metadata |73| Deserial | `->__wakeup()`, `->__destruct()`, `->__toString()`, `->__call()` | magic method defined on reachable class = POP gadget |74| File incl | `include($x)`, `include_once($x)` | LFI → RCE if attacker controls content (log poisoning, `/proc/self/environ`, phar://) |75| File incl | `require($x)`, `require_once($x)` | same as include |76| Command | `system($x)`, `exec($x)`, `passthru($x)` | record if `escapeshellarg`/`escapeshellcmd` used on `$x` |77| Command | `shell_exec($x)`, `popen($x, ...)`, `proc_open($x, ...)` | same |78| Command | `pcntl_exec($path, $args)` | args array — individual escape not needed but path is |79| SQL | `mysql_query($x)`, `mysqli_query($db, $x)`, `$mysqli->query($x)` | is `$x` concatenated from user input? |80| SQL | `pg_query($db, $x)` | same |81| SQL | `$pdo->query($x)`, `$pdo->exec($x)` | if query uses concat — SQLi. `prepare()` with `?` / `:name` is safe IF bindParam/execute receives separate args |82| SQL | `$pdo->prepare(...)` followed by `bindParam` with `PDO::PARAM_STR` but query concatenates table/column names | prepared statements DON'T protect identifiers — concat of table/column = still SQLi |83| ORM raw | `DB::raw($x)`, `DB::select($raw, ...)`, `Eloquent::whereRaw($x)` (Laravel) | is `$x` user input? bindings array separate? |84| ORM raw | `$wpdb->query($x)` without `$wpdb->prepare()` (WordPress) | |85| ORM raw | `$this->db->query($x)` (CodeIgniter) | |86| XSS | `echo $x`, `print $x`, `<?= $x ?>`, `printf($fmt, $x)` | escaped? (htmlspecialchars with ENT_QUOTES & correct charset? context-correct: HTML body vs attribute vs JS vs URL vs CSS?) |87| XSS | `echo "<img src=$x>"` inside `onclick=` etc. | attribute context — quoting matters |88| XSS | Template: `{{ $x }}` Blade (auto-escaped, OK) vs `{!! $x !!}` Blade (raw — DANGEROUS) | |89| XSS | Twig `{{ x\|raw }}`, `{% autoescape false %}` | |90| Open redirect | `header("Location: $x")` | domain check? |91| Header inj | `header($x)` or `header("X-Custom: $x")` where `$x` contains `\r\n` | CRLF → response splitting |92| File read | `file_get_contents($path)`, `fopen($path, 'r')`, `readfile($path)`, `file($path)`, `show_source($path)`, `highlight_file($path)` | path validation? `realpath` + basedir check? `..` filter? null byte (`\0`)? |93| File write | `file_put_contents($path, $content)`, `fwrite($fp, ...)`, `copy($src, $dst)`, `move_uploaded_file($tmp, $dst)` | where is `$dst` chosen? extension check? MIME on server? |94| File delete | `unlink($path)`, `rmdir($path)` | path traversal = arbitrary delete |95| Dir | `scandir($path)`, `glob($path)`, `opendir($path)` | info disclosure |96| SSRF | `curl_exec($ch)` with user URL, `file_get_contents("http...")`, `fsockopen($host, $port)`, `stream_socket_client`, `get_headers($url)` | URL allowlist? followed redirects? gopher/file scheme blocked? |97| XXE | `simplexml_load_string($xml)`, `simplexml_load_file($xml)`, `DOMDocument::loadXML`, `SoapClient::__doRequest` | `libxml_disable_entity_loader(true)` called? `LIBXML_NOENT` flag? PHP 8.0+ default is safer |98| Dynamic call | `call_user_func($fn, ...)`, `call_user_func_array($fn, ...)`, `$obj->$method()`, `$fn(...)` where `$fn` is variable | attacker picks function/method → RCE |99| Dynamic class | `new $cls($args)`, `$cls::method()` | attacker-picked class instantiation — constructor RCE via phpggc-like chain |100| Type juggle | `==` (loose compare) on auth token / hash / secret | `"0e1234..." == "0e5678..."` both truthy (both parse to `0e` scientific notation → 0) |101| Type juggle | `in_array($x, $arr)` without 3rd arg `true` | loose compare, same bypass |102| Type juggle | `strcmp($a, $b)` with array arg (PHP < 8) | returns NULL, loosely-equal to 0 → "match" |103| Auth bypass | `md5($pass) == $stored`, `sha1(...)` | weak + type juggling |104| Auth bypass | `hash_equals` missing → timing side-channel | |105| LDAP | `ldap_search($ds, $base, $filter)` | filter concat = LDAP injection |106| XPath | `$xml->xpath($x)`, `DOMXPath::query($x)` | concat = XPath injection |107| Mail | `mail($to, $subj, $body, $headers)` | header injection via `\r\n` in `$to`/`$subj`/`$headers` |108| Session fixation | `session_id($x)` where `$x` from user | attacker sets victim session id |109| Regex DoS | `preg_match($pattern, $subject)` where `$pattern` is user | ReDoS + `/e` (if PHP < 7) |110| Template inj | `Twig\Environment::createTemplate($user)`, Smarty `display("string:$user")`, Blade `compile($user)` | SSTI |111| Mass assignment | `$user->fill($_POST)`, `$user->forceFill(...)`, `(new User)->guard([])->fill(...)` | `$fillable` vs `$guarded` mis-config → privilege escalation |112| Path concat in stream wrappers | `"php://filter/resource=$file"`, `"zip://$x"`, `"data://text/plain,$x"` | filter chain for arbitrary read/write / code injection via `convert.base64-decode` |113| Insecure rand | `rand()`, `mt_rand()` for tokens/CSRF/reset | use `random_bytes` / `random_int` |114| Debug | `var_dump`, `print_r`, `phpinfo()`, `debug_print_backtrace`, `xdebug_*` in production | info disclosure if reachable |115116PHP "guard" to record (ALWAYS note if present/absent and the exact line):117- `escapeshellarg`, `escapeshellcmd` (cmd injection — note that `escapeshellcmd` is NOT sufficient for args; `escapeshellarg` wraps each arg)118- `htmlspecialchars($x, ENT_QUOTES, 'UTF-8')` (note flags and charset — missing flags = attribute context bypass)119- `htmlentities` (similar)120- `strip_tags` (weak — context-dependent)121- `addslashes` (NOT sufficient for SQL — use prepared statements)122- `mysqli_real_escape_string` (only valid with correct connection charset; `GBK` encoding attacks)123- `PDO::prepare` + `bindParam`/`bindValue`/`execute([...])` (safe for values, NOT identifiers)124- `basename($path)`, `realpath($path)`, `pathinfo()` (path normalization — still needs allowlist)125- `filter_var($url, FILTER_VALIDATE_URL)` (weak — `http://allowed.com@evil.com` passes)126- `parse_url` checks (often bypassable)127- `is_numeric` (allows `0x1A`, `1e10`, `+1`, `-1`, leading whitespace — NOT a security filter)128- `ctype_digit` (stricter — only `0-9`)129- `hash_equals` (timing-safe compare)130- `password_verify` (proper — uses `hash_equals` internally)131- Framework middleware: Laravel `CSRF`, `auth`, `throttle`; Symfony `security.yaml` firewall; CodeIgniter CSRF token — note if present on the route but don't assume it works132133## For Each Dangerous Operation, Record1341351. **Operation**: exact function/pattern and line number1362. **Category**: memory / integer / control-flow / concurrency / language-specific1373. **Operands**: what variables feed into it1384. **Guard**: what check (if any) protects this operation? Line number of the check.1395. **Neighbors**: what operations happen immediately before/after (±5 lines)?140141## Output142143Write to `sast-work/<file_hash>-dangers.json`:144```json145{146 "file": "src/codec/h264_slice.c",147 "dangerous_ops": [148 {149 "operation": "memset(slice_table, -1, sizeof(slice_table))",150 "line": 83,151 "category": "memory",152 "operands": {"dst": "slice_table (uint16_t[])", "value": "-1 = 0xFF per byte = 65535 per entry", "size": "frame_width * frame_height / 256"},153 "guard": "none — unconditional initialization",154 "neighbors": "line 85: slice_count = 0 (int32_t)",155 "note": "memset -1 on uint16 creates sentinel value 65535. If slice_count (int32) reaches 65535, it collides with sentinel."156 },157 {158 "operation": "slice_table[mb_pos] = slice_count",159 "line": 147,160 "category": "integer",161 "operands": {"dst": "uint16_t", "src": "int32_t (truncation)"},162 "guard": "none — slice_count not bounded",163 "neighbors": "line 149: comparison uses slice_table[neighbor] == slice_table[mb_pos]"164 }165 ]166}167```168169## Rules170171- **Be exhaustive.** List every matching pattern. A missed dangerous op = a missed vulnerability later.172- **Record the guard even if it looks correct.** The gap-analyzer will decide if it's sufficient.173- **The "note" field is optional** — use it only when you see something obviously interesting (like the sentinel collision above). But keep it brief and factual.174- Do NOT conclude anything is vulnerable. You are a cataloger, not an analyst.175176## Brain Integration177Check brain for prior analysis. Skip if already mapped and file unchanged.178179## Top-Tier Operator Standard180181Danger mapping is a catalog of risky operations plus their guards.182183- Record sink, operation type, arguments, enclosing function, caller hint, and nearby validation.184- Include "boring" guards: length checks, type checks, auth checks, escaping, canonicalization, feature flags, and error handling.185- Do not infer vulnerability. Mark uncertainty and let flow/gap agents reason over it.186- Prioritize sinks that transform attacker-controlled data into memory access, command execution, file access, network access, template rendering, deserialization, SQL, auth decisions, or state mutation.187- Preserve line numbers and exact identifiers so later agents can trace without rereading the whole file.
h-mmer