Skip to content

configuration.md

Latest commit

 

History

History
322 lines (254 loc) · 11.5 KB

configuration.md

File metadata and controls

322 lines (254 loc) · 11.5 KB
Raw

Configuration

Project Settings

[project]
name = "my-project"
path = "/path/to/target"
languages = ["c", "cpp", "python"]

Scanner Performance & Phase Flags

The [scanner.performance] section controls incremental scanning and which optional analysis phases run. Each phase flag defaults to a safe value; side-effect-heavy phases (auto-patching, PoC compilation) are opt-in.

[scanner.performance]
# Skip unchanged files based on SHA256 hash comparison (hashes persisted to output dir)
enable_incremental_scan = false
# Maximum number of parallel tasks for scanning operations
max_parallel_tasks = 4
# Enable LLM response caching to avoid redundant API calls
enable_llm_cache = false
# Enable file filtering to reduce false positives
enable_file_filtering = true

# --- Phase enable flags ---
# Each flag controls whether a specific analysis phase runs during the scan.

# Threat modeling using STRIDE analysis (adds LLM-based threat identification)
enable_threat_modeling = true
# Root-cause deduplication (collapses findings that share the same root cause)
enable_root_cause_dedup = true
# Multi-verifier cross-checking (runs additional LLM verification passes)
enable_multi_verifier = true
# Auto-patching (generates and validates fix patches in a staging worktree)
# Writes code files and runs git commands — opt-in for safety
enable_auto_patching = false
# PoC compilation (compiles proof-of-concept exploits to verify findings)
# Spawns external compilers — opt-in for safety
enable_poc_compilation = false
# Confidence refinement (re-calibrates finding confidence based on multi-source/cross-file signals)
enable_confidence_refinement = true
# CVE bootstrap (enriches findings with CVE data from external sources)
enable_cve_bootstrap = true
# Variant search (searches for variant instances of the same vulnerability pattern)
enable_variant_search = true
Flag Default Side effects
enable_threat_modeling true None (read-only analysis)
enable_root_cause_dedup true None
enable_multi_verifier true Additional LLM API calls
enable_auto_patching false Writes code files, runs git commands in a staging worktree
enable_poc_compilation false Spawns external compilers
enable_confidence_refinement true None
enable_cve_bootstrap true External network requests to NVD/CISA
enable_variant_search true Additional LLM API calls

See docs/architecture.md for the full 20-phase pipeline description.

LLM Configuration

BACO supports single or multiple models per phase. When multiple models are configured, they are used in round-robin fashion to distribute load across different models/providers.

Detailed error logging: When LLM requests fail, BACO reports the HTTP status code, error type (timeout, connection, request, body, decode), and the actual URL for easier debugging.

Single model:

[llm.phases.discovery]
base_url = "https://api.mistral.ai/v1"
api_key = "${MISTRAL_API_KEY}"  # or set env var
model = "mistral-small"

Multiple models:

[llm.phases.discovery]
base_url = "https://api.mistral.ai/v1"
api_key = "${MISTRAL_API_KEY}"
# 'models' takes precedence over 'model' if both are present
models = ["mistral-small", "mistral-medium", "codestral-latest"]

[llm.phases.verification]
base_url = "https://api.qwen.ai/v1"
api_key = "${QWEN_API_KEY}"
model = "qwen35"  # single model

[llm.phases.aggregation]
base_url = "https://api.openai.com/v1"
api_key = "${OPENAI_API_KEY}"
models = ["gpt-4o", "gpt-4o-mini"]  # multiple models for distributed load

Note: The models array takes precedence over model if both are present. Models are selected in round-robin fashion to distribute load across different providers.

Agent Mode

BACO has two distinct agent modes:

1. Discovery Agent (agent.enabled = true)

When enabled, the LLM Discovery phase reads source files directly before analyzing findings:

[agent]
enabled = true
max_turns = 10           # Max conversation turns with tools
tool_timeout_secs = 60   # Timeout for tool execution
keep_artifacts = false   # Keep generated test files

Benefits:

  • LLM reads actual source code before enriching findings
  • Uses tools (file_read, pattern_search) for deeper analysis
  • Provides more accurate vulnerability descriptions with context

2. SecurityAgent Verification (Phase 7)

A separate verification phase that uses an embedded security agent with tools to prove or disprove findings:

  • file_read: Examine vulnerable code in context
  • pattern_search: Look for related vulnerability patterns
  • file_write: Create proof-of-concept test cases
  • run_test: Execute tests to verify exploitability

The agent automatically removes false positives when tests pass, reducing noise in the final report. This phase runs after LLM Verification and before Ticket Cross-Reference.

Prompt Customization

BACO uses prompt templates for each phase loaded from markdown files at runtime. You can override these via configuration:

Default prompts are stored in prompts/phases/ as markdown files:

  • prompts/phases/indexing.md
  • prompts/phases/semgrep.md
  • prompts/phases/llm_static_analysis.md
  • prompts/phases/llm_discovery.md
  • prompts/phases/llm_verification.md
  • prompts/phases/ticket_crossref.md
  • prompts/phases/git_analysis.md
  • prompts/phases/cross_file_analysis.md
  • prompts/phases/confidence_scoring.md
  • prompts/phases/ai_aggregation.md
  • prompts/phases/reporting.md

View the full prompt templates on GitHub to understand default behavior.

Inline override in config.toml:

[llm.phases.prompt_overrides.phases]
llm_static_analysis = """Analyze this %%LANGUAGE%% code for security vulnerabilities.
Focus on: memory safety, injection risks, and insecure API usage.

File: %%FILE_PATH%%
Code:
%%CODE_CONTENT%%
"""

llm_discovery = """Given this finding, determine if it's a true vulnerability:
Title: %%FINDING_TITLE%%
Location: %%FILE_PATH%%:%%LINE_NUMBER%%
Description: %%VULNERABILITY_DESCRIPTION%%
"""

Available template variables:

  • %%PROJECT_PATH%% - Target project path
  • %%FILE_EXTENSIONS%% - Detected file extensions
  • %%LANGUAGES%% - Target languages
  • %%CODE_CONTENT%% - Code snippet being analyzed
  • %%LANGUAGE%% - Programming language of the file
  • %%FILE_PATH%% - File path
  • %%LINE_RANGE%% - Line numbers
  • %%FINDING_TITLE%% - Vulnerability title
  • %%VULNERABILITY_DESCRIPTION%% - Description text
  • %%FINDINGS_COUNT%% - Total findings count
  • %%SCAN_DATE%% - Scan date

Prompts are validated (max 10,000 characters, no null bytes) before use.

Ticket Systems

[[tickets.systems]]
type = "github"
url = "https://api.github.com"
credentials.token = "${GITHUB_TOKEN}"

Output Formats

  • findings.json: Complete vulnerability data with all 16 fields
  • report.html: Visual report with severity colors, code snippets, AI summary
  • findings.json: Complete vulnerability data with all 16 fields
  • report.html: Visual report with severity colors, code snippets, AI summary
  • report.sarif: SARIF format for CI/CD integration

Paper-Integration Research Flags

The following flags enable experimental research-backed analysis augmentations. All default to disabled. See todo.md for full implementation details.

VulTriage (P1) — arXiv:2605.09461

Triple-path context augmentation. Prepends control path (AST/CFG/DFG), knowledge path (CWE pattern RAG), and semantic path (function summary) to the LLM prompt before the vulnerability judgement.

Field Type Default Description
enabled bool false Enable triple-path augmentation
control_path bool true Include AST/CFG/DFG verbalisation
knowledge_path bool true Include CWE pattern RAG
semantic_path bool true Include function summary 
[vultriage]
enabled = false
control_path = true
knowledge_path = true
semantic_path = true

VulnLLM-R Policy Sampling (P2.2) — arXiv:2605.09461

Policy-based CWE generation. Queries the LLM N times to build a candidate set, then a final call picks one label. Increases LLM cost ~5x.

Field Type Default Description
enabled bool false Enable policy sampling
samples int 4 Number of candidate samples 
[policy_sampling]
enabled = false
samples = 4

VulnLLM-R Agent Scaffold (P2.5) — arXiv:2605.09461

Builds 3-path call-graph context + function-lookup tool per target for agent-assisted analysis.

Field Type Default Description
enabled bool false Enable agent scaffold
max_rounds int 5 Max agent conversation rounds
paths_per_target int 3 Call-graph paths per target 
[agent_scaffold]
enabled = false
max_rounds = 5
paths_per_target = 3

Truncated Generation (P2.1) — arXiv:2605.09461

Caps reasoning tokens before forcing the final answer. Configured in the [llm] section.

Field Type Default Description
max_reasoning_tokens int 2048 Max tokens for reasoning phase 
[llm]
max_reasoning_tokens = 2048

MoCQ Neuro-Symbolic Rule Synthesis (P3) — arXiv:2605.13918

RuleSynthesis 2.0: LLM proposes patterns in a DSL → symbolic validator gives feedback → iterative loop. Extends the [scanner.rulesynth] section.

Field Type Default Description
mocq_mode bool false Enable MoCQ neuro-symbolic mode
max_iterations int 5 Max synthesis iterations
corpus_path str "tests/fixtures/" Path to pattern corpus 
[scanner.rulesynth]
mocq_mode = false
max_iterations = 5
corpus_path = "tests/fixtures/"

PacVD Primitive-API Abstraction (P4) — arXiv:2605.07785

Appends callee abstraction at one of four granularity levels to the LLM prompt. Level 1 = fuzzy branches only; Level 4 = concrete branches + key variables.

Field Type Default Description
enabled bool false Enable primitive-API abstraction
level int 2 Abstraction granularity (1-4)
auto_level bool false Auto-select optimal level 
[pacvd]
enabled = false
level = 2
auto_level = false

AgentFlow Multi-Agent Harness Synthesis (P5) — arXiv:2605.11835

Represents the harness as a typed graph DSL with a search loop. Most invasive integration — static harness only until P5.5.

Field Type Default Description
enabled bool false Enable AgentFlow harness
max_iterations int 10 Max synthesis iterations
requires_instrumented_target bool false Require instrumented target 
[agent_flow]
enabled = false
max_iterations = 10
requires_instrumented_target = false