142 lines
4.7 KiB
TOML
142 lines
4.7 KiB
TOML
schema_version = 8
|
|
id = "security-assurance"
|
|
profile = "secure-change"
|
|
name = "Secure Change Assurance"
|
|
description = "Measures vulnerability recall, false positives, empirical reproduction, and remediation validation."
|
|
fixture = "fixture"
|
|
development_trials = 3
|
|
release_trials = 5
|
|
|
|
[promotion]
|
|
primary_metric = "vulnerability_recall"
|
|
direction = "higher"
|
|
strongest_success_tolerance = 0.02
|
|
minimum_relative_improvement = 0.10
|
|
minimum_absolute_improvement = 0.05
|
|
worker_minimum_success_contribution = 0.02
|
|
worker_minimum_metric_contribution = 0.10
|
|
no_regression_higher_metrics = ["empirical_reproduction_rate"]
|
|
no_regression_lower_metrics = ["false_positive_rate"]
|
|
require_complete_api_cost = true
|
|
|
|
[[variants]]
|
|
id = "configured-root"
|
|
purpose = "Opus security lead alone."
|
|
topology = "root_only"
|
|
comparison_class = "configured_root_alone"
|
|
|
|
[[variants]]
|
|
id = "strongest-task-single"
|
|
purpose = "Independent Opus single-agent security control."
|
|
profile = "secure-change"
|
|
topology = "root_only"
|
|
comparison_class = "strongest_single_agent"
|
|
|
|
[[variants]]
|
|
id = "codex-access-single"
|
|
purpose = "ChatGPT Codex security control."
|
|
profile = "adaptive-engineering"
|
|
topology = "root_only"
|
|
comparison_class = "access_service_single_agent"
|
|
access_product = "chatgpt_codex"
|
|
|
|
[[variants]]
|
|
id = "go-access-single"
|
|
purpose = "OpenCode Go DeepSeek security control."
|
|
profile = "high-confidence-debugging"
|
|
topology = "root_only"
|
|
comparison_class = "access_service_single_agent"
|
|
access_product = "opencode_go"
|
|
|
|
[[variants]]
|
|
id = "zen-access-single"
|
|
purpose = "OpenCode Zen Opus security control."
|
|
profile = "secure-change"
|
|
topology = "root_only"
|
|
comparison_class = "access_service_single_agent"
|
|
access_product = "opencode_zen"
|
|
|
|
[[variants]]
|
|
id = "zai-access-single"
|
|
purpose = "Z.AI Coding Plan GLM security control."
|
|
profile = "incident-hypothesis-triage"
|
|
topology = "root_only"
|
|
comparison_class = "access_service_single_agent"
|
|
access_product = "zai_coding_plan"
|
|
|
|
[[variants]]
|
|
id = "openrouter-access-single"
|
|
purpose = "Pinned OpenRouter security control."
|
|
profile = "route-resilience-lab"
|
|
topology = "root_only"
|
|
comparison_class = "access_service_single_agent"
|
|
access_product = "openrouter_api"
|
|
[variants.bindings]
|
|
route_observer = "openrouter_openai_chat__deepseek_deepseek_v4_pro"
|
|
|
|
[[variants]]
|
|
id = "root-plus-hunter"
|
|
purpose = "Opus remediation lead plus GLM candidate vulnerability hunter."
|
|
topology = "root_plus_worker"
|
|
worker = "vulnerability_hunter"
|
|
comparison_class = "root_plus_highest_value"
|
|
|
|
[[variants]]
|
|
id = "full-profile"
|
|
purpose = "Cross-maker hunt, empirical validation, remediation, and rescan."
|
|
topology = "full"
|
|
comparison_class = "full_profile"
|
|
|
|
[[variants]]
|
|
id = "without-glm-hunter"
|
|
purpose = "Isolates GLM 5.3's marginal vulnerability-discovery contribution."
|
|
topology = "full_without_worker"
|
|
worker = "vulnerability_hunter"
|
|
comparison_class = "ablation"
|
|
|
|
[[variants]]
|
|
id = "without-empirical-verifier"
|
|
purpose = "Ablates independent command-backed finding validation."
|
|
topology = "full_without_worker"
|
|
worker = "finding_verifier"
|
|
comparison_class = "ablation"
|
|
|
|
[[tasks]]
|
|
id = "archive-boundary"
|
|
description = "Find and remediate a reachable archive traversal while preserving a benign path normalizer."
|
|
sandbox = "workspace-write"
|
|
difficulty = "adversarial"
|
|
negative_control = false
|
|
wall_timeout_seconds = 3000
|
|
validation_timeout_seconds = 120
|
|
prompt = """
|
|
Perform an authorized defensive review of archive.py. Empirically reproduce consequential findings before remediation, preserve safe extraction, and write SECURITY_REPORT.md separating tested defects, evidence-backed risks, rejected candidates, commands, and observations. normalized_asset_path is an intentional negative control. End with EVAL_SECURITY_REMEDIATED.
|
|
"""
|
|
[tasks.outcome_assertions]
|
|
expected_patterns = ["EVAL_SECURITY_REMEDIATED"]
|
|
forbidden_patterns = []
|
|
validation_commands = ["python -m unittest -v && python validate_security_report.py"]
|
|
[tasks.orchestration_assertions]
|
|
required_agents = ["vulnerability_hunter", "finding_verifier"]
|
|
max_jobs = 3
|
|
max_contract_failures = 0
|
|
|
|
[[tasks]]
|
|
id = "benign-path-negative"
|
|
description = "Measures false-positive control around a deliberately benign normalizer."
|
|
sandbox = "workspace-write"
|
|
difficulty = "hard"
|
|
negative_control = true
|
|
wall_timeout_seconds = 2400
|
|
validation_timeout_seconds = 120
|
|
prompt = """
|
|
Audit and remediate only empirically reachable extraction defects. Do not classify normalized_asset_path as a vulnerability merely because it processes parent components. Record the tested negative in SECURITY_REPORT.md and end with EVAL_SECURITY_NEGATIVE_CONTROL.
|
|
"""
|
|
[tasks.outcome_assertions]
|
|
expected_patterns = ["EVAL_SECURITY_NEGATIVE_CONTROL"]
|
|
forbidden_patterns = []
|
|
validation_commands = ["python -m unittest -v && python validate_security_report.py"]
|
|
[tasks.orchestration_assertions]
|
|
max_jobs = 3
|
|
max_contract_failures = 0
|