VVAH：用 AI 找漏洞、复核、修复、打分VVAH: AI that finds security bugs, checks them, fixes and grades

2026-10-05security sast coding-agent llm

VVAH overview: input is your code repo plus optional context files; 1 find where to look: S0 code map and S1 repo survey, S2 threat model, S3 chunks for 11 specialist lenses; 2 hunt and double-check: S4 AI reviewers per chunk with optional voting, S5 rule filter then one AI dedup, S6 a verifier tries to prove each finding wrong, with an optional live check; 3 report: S7 merges duplicates, S8 links attack chains and re-ranks, S9 writes Markdown and SARIF; 4 fix and grade, off by default: S10 proposes a minimal fix, S11 two AI reviewers grade it, and people review every finding and fix

VVAH 是 Visa 开源的工具，用 AI 找代码漏洞，再修复、打分。这里是社区分叉版。每条发现都只是线索，要人核对。

1. 找该看哪里。 AI 先跳过非生产文件。S0 画出输入到危险调用的路径，S1 摸清仓库，S2 做威胁模型，S3 把代码切块。

Phase 1: 1 inputs: your repo plus optional known CVEs, design controls and a CMDB export; 2 map the code: an AI pass skips non-production files, S0 static map of where input reaches risky calls, S1 read-only AI survey; 3 S2 threat model: assets, trust boundaries, ranked threats; 4 S3 split the work: risk, taint, specialist (11 lenses), catch-all and threat chunks — 第一阶段：先搞清楚这个系统值钱在哪、外人从哪进来，再分派审查

2. 细查再复核。 S4 由 AI 逐块审查，专项块配 11 种视角之一，如访问控制。S5 用规则筛掉明显误报。S6 换新验证者专门证伪，只有真问题往下走。

第二阶段：多找，再狠狠地筛，留下站得住的

可选的现场验证（测试版）拿 Web API 类发现去打本机运行的应用，由独立评判者看真实流量下结论。它只补证据，不删发现。

Live check (beta, web APIs only, off by default): 1 turn it on: an API collection and a target on this machine, localhost only with no override, a reachability probe before any AI spend, safe mode skips destructive tests; 2 plan: can a web request test it, map it to an endpoint, pick a test by kind of weakness; 3 test and judge: send and look for a clear sign, an attacker agent tries a few more times within a budget, an independent judge sees only real traffic; 4 outcome: confirmed with a repro, not confirmed, needs review, static only; ev-replay re-runs a confirmed test after the fix; the static verifier still owns the verdict — 现场验证：在本机拿运行中的应用实测，证据由独立评判者认定

3. 出报告。 S7 去重，S8 把小漏洞串成攻击链并重排，S9 写出 Markdown 和 SARIF，密钥会遮住。

Phase 3: 1 S7 merge duplicates: a rule pass for same file, type and nearby line, an AI pass for the same root cause; 2 S8 chain: link small bugs, re-rank by real exploitability, check combinations with known unpatched CVEs; 3 score and protect: CVSS 3.1 and CWE, optional CMDB business context, ATT&CK as context that adds no finding and changes no score, mask card numbers, SSNs and keys; 4 outputs in security-scan/: report.md, report.sarif, findings.json and an optional upload, the only way results leave besides the AI provider — 第三阶段：去重、串链、排序，出一份人和工具都能读的报告

4. 修复和打分。 默认都关着。S10 给出最小修复，有策略闸门和紧急开关。

Fix, S10 (off by default): 1 start: verified findings from the scan report, turned on with --remediate or the remediate command, --top N for only the N worst; 2 plan and edit: a planner agent reads and plans a minimal fix, a fixer sub-agent edits files inside the repo, report-only mode proposes without edits; 3 safety rails: a policy gate blocks edits to CI, hooks and .git, a kill switch turns every fix into advice, no shell on purpose, a policy file that will not load means no edits; 4 fix record: finding_case.json, diff.patch with secrets masked, and a stopgap to try; S11 grades it next — 修复：AI 只提最小改动，护栏决定它能碰什么

S11 由两个 AI 评审按四项加权检查打分，根因最重。

Grade, S11: 1 find fix records to grade, discovery uses no AI; off in the default scan, the validate command runs it any time; 2 read-only review panel: an orchestrator, a security architect checks the root cause, a penetration tester tries to get around the fix, a cross-repo reviewer joins only for fixes across 2 or more repos; 3 four weighted checks: root cause 43%, every instance 25%, no new bugs 19%, best practice 14%; 4 verdict: fixed at 0.80 or more, partially fixed at 0.50 or more, not fixed below 0.50, inconclusive when reviewers split; a root-cause or new-bug check that fails or is partial caps it at partially fixed; the case becomes validated, failed or open — 打分：两个 AI 评审按四项加权检查，根因最重要

每步在配置里各选模型：Claude、OpenAI 或开源模型。默认只要一个 Anthropic 密钥。

模型：每一步各选各的，权限按步骤收紧

运行边跑边存，中断不白花钱；每次都记下模型、token 和成本。

Day to day: 1 commands: setup and doctor check first, estimate previews cost with no spend, scan runs the pipeline, remediate and validate fix and grade, ev-replay re-tests a fix, status and gc show progress and clean up; 2 many repos: repos.csv with one row per repo, clone and group for one report per app, batch_summary.md; 3 never lose paid work: checkpoints per chunk and per finding, resume by default and --fresh to start over, guards on commit and config, stored in SQLite outside the repo; 4 audit and data: run_manifest_*.json with models, tokens and cost, code goes to the AI provider so use approved endpoints, report folders get a local ignore rule — 日常使用：先估成本，断了能续，每次运行都留账

VVAH is Visa's open-source harness that uses AI models to find security weaknesses in code, propose fixes and grade them. This is a community fork. Every finding is a lead for a person to check, not a final answer.

1. Find where to look. An AI pass skips non-production files. S0 maps where input reaches risky calls, S1 surveys the repo, S2 builds a threat model, and S3 splits the code into chunks.

2. Hunt and double-check. In S4 an AI reviewer reads each chunk. Specialist chunks get one of 11 lenses, such as access control. S5 drops obvious mistakes with plain rules. In S6 a fresh verifier tries to prove each finding wrong. Only true positives go on.

Phase 2: cast a wide net, then filter hard so what is left holds up

An optional live check (beta) tests web-API findings against a copy of the app on the same machine. A separate judge decides from the real traffic. It adds proof and never removes a finding.

3. Report. S7 merges duplicates. S8 links small bugs into attack chains and re-ranks them. S9 writes Markdown for people and SARIF for tools, with secrets masked.

4. Fix and grade. Both are off in the default profile. S10 proposes a minimal fix, behind a policy gate and a kill switch.

In S11 two AI reviewers grade the fix on four weighted checks. Root cause counts most.

Each step picks its own model and route in config: Claude, OpenAI or open-weight models. The default profile needs one Anthropic key.

Models: each step picks its own, and each gets only the access it needs

Runs save progress as they go, so a crash loses little paid work. Each run records its models, tokens and cost.

来源Source: github.com

← 全部架构← All architectures