Skip to content

Commit dc324e7

Browse files
committed
fix: sanitize LLM finding metadata
Assisted-by: Codex (GPT-5) Signed-off-by: Efe Gökdemir <[email protected]>
1 parent 9902be4 commit dc324e7

5 files changed

Lines changed: 65 additions & 5 deletions

File tree

‎src/skillspector/llm_analyzer_base.py‎

Lines changed: 8 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -70,7 +70,6 @@
7070
from skillspector.logging_config import get_logger
7171
from skillspector.model_info import get_max_input_tokens
7272
from skillspector.models import Finding
73-
from skillspector.nodes.analyzers.pattern_defaults import get_category, get_pattern_name
7473

7574
logger = get_logger(__name__)
7675

@@ -615,17 +614,22 @@ def _normalize_confidence(cls, v: object) -> float:
615614

616615
def to_finding(self, file: str) -> Finding:
617616
"""Convert to a :class:`Finding` for the graph state."""
617+
# Keep the core LLM module independent from analyzer discovery. The
618+
# discovery package imports this module while loading analyzers.
619+
from skillspector.nodes.analyzers.pattern_defaults import get_category, get_pattern_name
620+
621+
message = self.message[:4096]
618622
return Finding(
619623
rule_id=self.rule_id,
620-
message=self.message,
624+
message=message,
621625
severity=self.severity,
622626
confidence=self.confidence,
623627
file=file,
624628
start_line=self.start_line,
625629
end_line=self.end_line,
626630
category=get_category(self.rule_id),
627-
pattern=self.message or get_pattern_name(self.rule_id),
628-
finding=self.message,
631+
pattern=message or get_pattern_name(self.rule_id),
632+
finding=message,
629633
explanation=self.explanation,
630634
remediation=self.remediation,
631635
)

‎src/skillspector/nodes/analyzers/pattern_defaults.py‎

Lines changed: 24 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -243,6 +243,18 @@ class PatternCategory(StrEnum):
243243
"DS2": PatternCategory.DESERIALIZATION.value,
244244
"DS3": PatternCategory.DESERIALIZATION.value,
245245
"DS4": PatternCategory.DESERIALIZATION.value,
246+
# Semantic LLM analyzers
247+
"SSD-1": PatternCategory.PROMPT_INJECTION.value,
248+
"SSD-2": PatternCategory.PROMPT_INJECTION.value,
249+
"SSD-3": PatternCategory.DATA_EXFILTRATION.value,
250+
"SSD-4": PatternCategory.PROMPT_INJECTION.value,
251+
"SDI-1": PatternCategory.EXCESSIVE_AGENCY.value,
252+
"SDI-2": PatternCategory.EXCESSIVE_AGENCY.value,
253+
"SDI-3": PatternCategory.EXCESSIVE_AGENCY.value,
254+
"SDI-4": PatternCategory.EXCESSIVE_AGENCY.value,
255+
"SQP-1": PatternCategory.TRIGGER_ABUSE.value,
256+
"SQP-2": PatternCategory.OUTPUT_HANDLING.value,
257+
"SQP-3": PatternCategory.OUTPUT_HANDLING.value,
246258
}
247259

248260
# Rule ID -> pattern display name (for report output)
@@ -334,6 +346,18 @@ class PatternCategory(StrEnum):
334346
"DS2": "Ruby Marshal Deserialization",
335347
"DS3": "Unsafe Ruby YAML Deserialization",
336348
"DS4": "Unsafe JavaScript Deserialization",
349+
# Semantic LLM analyzers
350+
"SSD-1": "Semantic Prompt Injection",
351+
"SSD-2": "Novel Attack Phrasing",
352+
"SSD-3": "Natural-Language Data Exfiltration",
353+
"SSD-4": "Narrative Deception",
354+
"SDI-1": "Description-Behavior Mismatch",
355+
"SDI-2": "Context-Inappropriate Capability",
356+
"SDI-3": "Permission Scope Creep",
357+
"SDI-4": "Intent-Code Divergence",
358+
"SQP-1": "Vague Trigger",
359+
"SQP-2": "Missing User Warning",
360+
"SQP-3": "Natural-Language Policy Violation",
337361
}
338362

339363
# Pattern-specific remediations (how to fix the issue)

‎src/skillspector/nodes/report.py‎

Lines changed: 2 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -102,6 +102,7 @@
102102
# Finding free-text fields that may carry scanned/LLM content.
103103
_SANITIZED_FIELDS = (
104104
"message",
105+
"pattern",
105106
"explanation",
106107
"remediation",
107108
"finding",
@@ -144,6 +145,7 @@ def clean_evidence(value: object) -> object:
144145
return replace(
145146
finding,
146147
message=clean(finding.message) or "",
148+
pattern=clean(finding.pattern),
147149
explanation=clean(finding.explanation),
148150
remediation=clean(finding.remediation),
149151
finding=clean(finding.finding),

‎tests/nodes/test_llm_analyzer_base.py‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -2556,7 +2556,7 @@ def test_to_finding_populates_report_metadata_for_semantic_rules(self) -> None:
25562556
start_line=3,
25572557
).to_finding("SKILL.md")
25582558

2559-
assert finding.category == "Security"
2559+
assert finding.category == "Data Exfiltration"
25602560
assert finding.pattern == "Natural-language credential exfiltration"
25612561
assert finding.finding == "Natural-language credential exfiltration"
25622562

‎tests/nodes/test_report_sanitizer.py‎

Lines changed: 30 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -21,6 +21,7 @@
2121

2222
import pytest
2323

24+
from skillspector.llm_analyzer_base import LLMFinding
2425
from skillspector.models import Finding
2526
from skillspector.nodes.report import _clean_text, _sanitize_finding, report
2627
from skillspector.state import SkillspectorState
@@ -36,6 +37,7 @@ def _dirty_finding() -> Finding:
3637
start_line=5,
3738
remediation="redact \x1b[1mnow\x1b[0m",
3839
context="line with \x07 bell and \x1b[0m reset",
40+
pattern="pattern \x1b[31mred\x1b[0m\x00",
3941
)
4042

4143

@@ -55,6 +57,7 @@ def test_sanitize_finding_cleans_text_fields_only() -> None:
5557
assert "leak" in cleaned.message and "here" in cleaned.message
5658
assert "\x1b" not in (cleaned.remediation or "")
5759
assert "\x07" not in (cleaned.context or "")
60+
assert "\x1b" not in (cleaned.pattern or "") and "\x00" not in (cleaned.pattern or "")
5861
# Non-text fields are unchanged.
5962
assert cleaned.rule_id == "E2"
6063
assert cleaned.start_line == 5
@@ -93,6 +96,7 @@ def test_report_redacts_url_credentials_from_every_finding_field(fmt: str, schem
9396
file="setup.sh",
9497
start_line=1,
9598
finding=url,
99+
pattern=url,
96100
explanation=url,
97101
remediation=url,
98102
context=url,
@@ -121,6 +125,32 @@ def test_report_redacts_url_credentials_from_every_finding_field(fmt: str, schem
121125
assert secret not in serialized_findings
122126

123127

128+
@pytest.mark.parametrize("fmt", ["json", "sarif"])
129+
def test_report_sanitizes_llm_message_copied_to_pattern(fmt: str) -> None:
130+
message = "review \x1b[31mhttps://user:[email protected]/?token=secret-token\x1b[0m"
131+
finding = LLMFinding(
132+
rule_id="E1",
133+
message=message,
134+
severity="HIGH",
135+
start_line=1,
136+
).to_finding("SKILL.md")
137+
state: SkillspectorState = {
138+
"filtered_findings": [finding],
139+
"component_metadata": [],
140+
"has_executable_scripts": False,
141+
"manifest": {},
142+
"skill_path": None,
143+
"output_format": fmt,
144+
}
145+
146+
result = report(state)
147+
rendered = result["report_body"]
148+
serialized_findings = json.dumps([item.to_dict() for item in result["filtered_findings"]])
149+
assert "\x1b" not in rendered
150+
assert "secret" not in rendered
151+
assert "secret" not in serialized_findings
152+
153+
124154
def test_nested_evidence_preserves_scalar_types_and_original_finding() -> None:
125155
scalar_values = [None, True, False, 42, 1.25]
126156
finding = _dirty_finding()

0 commit comments

Comments
 (0)