惯性聚合 高效追踪和阅读你感兴趣的博客、新闻、科技资讯
阅读原文 在惯性聚合中打开

推荐订阅源

D
DataBreaches.Net
IT之家
IT之家
The Cloudflare Blog
Apple Machine Learning Research
Apple Machine Learning Research
WordPress大学
WordPress大学
N
Netflix TechBlog - Medium
阮一峰的网络日志
阮一峰的网络日志
P
Proofpoint News Feed
L
LangChain Blog
博客园 - Franky
美团技术团队
J
Java Code Geeks
Microsoft Security Blog
Microsoft Security Blog
博客园 - 叶小钗
小众软件
小众软件
Y
Y Combinator Blog
B
Blog RSS Feed
Cyber Security Advisories - MS-ISAC
Cyber Security Advisories - MS-ISAC
D
Docker
Hugging Face - Blog
Hugging Face - Blog
Jina AI
Jina AI
罗磊的独立博客
大猫的无限游戏
大猫的无限游戏
Vercel News
Vercel News

Hacker News - Newest: "AI"

AI can't read an investor deck AI as an attorney? Student uses ChatGPT, Gemini to sue UW over alleged racial discrimination Hacking MCP Servers in AI Systems – The Rug Pull: Tool Changes After Approval GitHub - MeepCastana/KubeezCut: Free Web based video editor Can AI judge journalism? A Thiel-backed startup says yes, even if it risks chilling whistleblowers Coming soon: 10 Things That Matter in AI Right Now DARPA built an AI to fact-check enemy weapons claims What explains heterogeneity in AI adoption? When AI Meets Muscle: Context-Aware Electrical Stimulation Promises a New Way to Guide Human Movements - Department of Computer Science AI Changed How We Build. It Did Not Change What Matters. Linux rules on using AI-generated code - Copilot is OK, but humans must take 'full responsibility for the… Meta spins up AI version of Mark Zuckerberg to engage with employees Code Mode: Let Your AI Write Programs, Not Just Call Tools | TanStack Blog GitHub - Delavalom/graft: Go framework for building AI agents. Type-safe tools, multi-provider (OpenAI, Anthropic, Gemini, Bedrock), zero vendor SDKs. India's TCS tops estimates, says new AI models did not dent services demand Gen Z's fading AI hype Strong feeling: we are in a folded AI reality GitHub - machinarii/total-recall-catalog: A reference catalog of latest knowledge retrieval, memory & RAG systems GitHub - mensfeld/code-on-incus: Give each AI agent its own isolated machine with root, Docker, and systemd. Active defense detects and stops threats automatically.. Quantization, LoRA, and the 8% Problem: Benchmarking Local LLMs for Production AI Iran war: We spoke to the man making Lego-style AI videos that experts say are powerful propaganda Powell, Bessent discussed Anthropic's Mythos AI cyber threat with major U.S. banks GitHub - immartian/bellamem: Persistent belief-graph memory for AI agents. Retrieves decisive context by importance — not recency, not RAG, not /compact. recursive-mode: The Repo-Native Operating System for AI Engineering After the attack on Sam Altman's home, will AI CEO's go on the offensive? The biggest advance in AI since the LLM Opus 4.6 vs GPT 5.4 One Prompt Unity World Generation Test “AI polls” are fake polls Client Challenge Can AI be a 'child of God'? Inside Anthropic's meeting with Christian leaders
GitHub - Diplomat-ai/diplomat-agent-ts: What can your Typ...
jguarnelli · 2026-05-28 · via Hacker News - Newest: "AI"

Discussions · Contributing · Changelog · Issues

npm version CI npm downloads Node 20+ License: Apache 2.0 OWASP Agentic

You shipped a TypeScript AI agent. Do you know every function it can call that writes to a database, sends an email, charges a card, or deletes data — and which ones have zero checks?

diplomat-agent-ts runs a static AST scan and tells you exactly that. Two dependencies. ~9 s on a 7,874-file TypeScript agent codebase (OpenClaw, M-series). ~30 s on slower x86 hardware without a tsconfig.json.

npm install -D @diplomat-ai/diplomat-agent-ts
npx diplomat-agent-ts scan .        # scan from project root
npx diplomat-agent-ts scan ./src    # or a specific subdirectory

Before diplomat-agent-ts: you can't see what your agent can do. After: every tool call is mapped, classified by side effect, and tagged with OWASP Agentic codes — diffable in PRs.

What it looks like

diplomat-agent-ts terminal output showing one unguarded chargeCustomer call with missing bounds, rate limit, and approval step (mapped to OWASP ASI-01, ASI-02, ASI-03, ASI-06), one unguarded deleteUserData call, and one confirmed sendWelcomeEmail call annotated with checked:ok.

Why this matters for AI agents

In a web app, a human clicks a button. The UI has validation, confirmation dialogs, rate limits per session.

In an agent, an LLM decides which functions to call, with what arguments, how many times. It doesn't know your business rules. It can loop, hallucinate arguments, or get prompt-injected.

Without guards in the code, there's nothing between the LLM's decision and the real-world consequence.

We scanned the OpenClaw agent codebase at pinned commit 49d9996d (7,874 TypeScript files, ~9 s on M-series, ~30 s on x86). 419 tool calls had real side effects. 332 of them (79%) had zero checks. Not a single one was confirmed.

What it detects

40+ patterns across 12 side-effect categories:

Category Examples (TS-native) Required guards
payment stripe.charges.create(), stripe.refunds.create() bounds, rate limit, approval
database_write Prisma .create() / .update(), Mongoose .save() input validation, rate limit
database_delete Prisma .delete(), Mongoose .deleteOne(), raw DELETE batch protection, confirmation
http_write axios.post(), fetch(POST), got.put() rate limit, retry bound
email nodemailer.sendMail(), resend.emails.send(), sgMail.send() rate limit
messaging twilio.messages.create(), slack.chat.postMessage() rate limit
agent_invocation agent.run(), graph.invoke(), Runner.run() input validation, approval
llm_call openai.chat.completions.create(), anthropic.messages.create()
publish s3.send(PutObjectCommand), client.publish() approval
dynamic_code eval(), new Function(), vm.runInNewContext() confirmation
file_delete fs.rm(), fs.unlink(), fs-extra.remove() confirmation
destructive execSync(), spawnSync(), execa() confirmation

What counts as a guard: input validation (Zod, Yup, class-validator), rate limiting (NestJS @Throttle, custom decorators), auth checks (NestJS guards, middleware), confirmation steps, idempotency keys, retry bounds. Full catalog in src/scanner/patterns.ts.

Quick start

# Scan from your project root (default: current directory)
diplomat-agent-ts scan .

# Or a specific subdirectory
diplomat-agent-ts scan ./src
diplomat-agent-ts scan ./packages

# Generate the toolcalls.yaml SBOM (commit this)
diplomat-agent-ts scan . --output-registry toolcalls.yaml

# Fail CI when new unguarded tool calls appear
diplomat-agent-ts scan . --fail-on-unchecked

# JSON output for IDE agents, automation, custom dashboards
diplomat-agent-ts scan . --format json

The scanner emits:

  • A coloured ANSI report to stdout (default)
  • toolcalls.yaml — a diff-stable registry of every detected tool call (with --output-registry)
  • JSON — snake_case field names, interoperable with the Python scanner (--format json)

Integrate everywhere

CI — block unguarded PRs

# .github/workflows/diplomat.yml
- name: Diplomat governance scan
  run: npx -y @diplomat-ai/diplomat-agent-ts scan . --fail-on-unchecked

Exit code 1 if any tool call has no_checks status. Exit 0 otherwise — even if partial_checks exist (they're warnings, not blockers).

Pre-commit hook

# .pre-commit-config.yaml
repos:
  - repo: local
    hooks:
      - id: diplomat-agent-ts
        name: diplomat governance scan
        entry: npx -y @diplomat-ai/diplomat-agent-ts scan . --fail-on-unchecked
        language: system
        pass_filenames: false

IDE — review what the copilot wrote

The scanner runs locally in under 10 seconds on typical agent codebases. Ask Claude Code, Copilot, or Cursor to run it after generating tool-calling code:

"Run diplomat-agent-ts scan . and fix any unguarded tool calls."

Note: AI agents (Claude Code, Copilot, Cursor) may summarize scan output inaccurately when the result is long. Always read the raw stdout — or pipe to a file with --output report.txt and read that.

Acknowledge a tool call

When a function is intentionally unguarded or protected outside the static-analysis scope, mark it inline:

// checked:ok — protected by middleware/approval.ts
export async function chargeCustomer(amount: number, customerId: string) {
  return stripe.charges.create({ amount, currency: "usd", customer: customerId });
}

diplomat:ok and canary:ok are accepted as aliases. The next scan moves the call to confirmed status with an empty missing_hints list. The annotation appears in the YAML registry so reviewers can audit why something was confirmed.

toolcalls.yaml — a behavioral SBOM

Like package-lock.json, but for what your agent can do, not what it depends on:

spec_version: "1.0"
language: typescript

summary:
  total: 12
  no_checks: 8
  partial_checks: 3
  confirmed: 1

tool_calls:
  - function: chargeCustomer
    file: src/payments.ts
    line: 42
    actions:
      - "return stripe.charges.create({ amount, currency, customer })"
    checks: []
    missing:
      - no bounds on amount
      - no rate limit
      - no idempotency key
    owasp: [ASI-01, ASI-02, ASI-03, ASI-06]

Commit it. Diff it in PRs. When your agent gains a new capability, the change shows up in review — before it ships.

Spec → docs/toolcalls-yaml-spec.md

OWASP Agentic Top 10 mapping

Each finding is tagged with one or more relevant codes from the OWASP Agentic Security Initiative Top 10. The v0.1.0 catalog covers the codes most directly tied to static side-effect detection (ASI-01, ASI-02, ASI-03, ASI-04, ASI-05, ASI-06, ASI-10). Codes that require runtime context (ASI-07 supply chain, ASI-08 misalignment, ASI-09 deception) are out of scope for static analysis — they are covered by diplomat-gate and diplomat.run at runtime.

Code Risk When it fires
ASI-01 Excessive Agency Side effect with no auth check
ASI-02 Tool Misuse Any side effect (baseline tag)
ASI-03 Privilege Compromise Payments, deletes, destructive ops with no confirmation
ASI-04 Resource Overload Agent invocations without bounds
ASI-05 Cascading Hallucination LLM call chained to side effect
ASI-06 Identity Spoofing Missing rate limit or retry bound
ASI-10 Overreliance Nested agent invocations

Full mapping in src/analyzer/owasp.ts.

Architecture

Internal architecture of diplomat-agent-ts in three lanes: SCANNER (patterns.ts, matcher.ts, ast-scanner.ts producing Tool[]), ANALYZER (checks.ts, owasp.ts producing ScanResult), REPORTER (registry.ts, json.ts, terminal.ts). Contribution hints at the bottom: add patterns in patterns.ts only, change matching logic in matcher.ts with tests, new output formats as new files in reporter/.

Three pure stages, no shared state. See CONTRIBUTING.md for full guidelines.

Benchmarks

Real codebases, real numbers.

Methodology. File counts are the number of .ts / .tsx files actually scanned after the scanner's built-in exclusions (node_modules/, dist/, build/, *.test.ts, *.spec.ts, *.d.ts). They can differ from a raw git ls-files count by a few percent. Runs are pinned to the commits below. Findings counts (tool_calls, no_checks, partial) reproduce exactly at those commits; raw file totals on main will drift over time.

Codebase (scope) Type TS files scanned Tool calls no_checks partial Pinned commit
OpenClaw (src/) Application 7,874 419 332 (79%) 87 49d9996d
Mastra (packages/) Framework 2,777 185 162 (88%) 23 38b87964
OpenAI Agents JS (packages/) Framework 426 33 31 (94%) 2 629d35af
OpenAI Agents JS (examples/) Examples 302 32 28 (88%) 4 629d35af

Run the benchmarks yourself:

# Application — OpenClaw (pinned commit for reproducibility)
git clone https://github.com/openclaw/openclaw /tmp/openclaw
cd /tmp/openclaw && git checkout 49d9996d && cd -
npx diplomat-agent-ts scan /tmp/openclaw/src

# Framework — Mastra
git clone --depth 1 https://github.com/mastra-ai/mastra /tmp/mastra
npx diplomat-agent-ts scan /tmp/mastra/packages

# Framework + Examples — OpenAI Agents JS
git clone --depth 1 https://github.com/openai/openai-agents-js /tmp/openai-agents
npx diplomat-agent-ts scan /tmp/openai-agents/packages
npx diplomat-agent-ts scan /tmp/openai-agents/examples

Output formats

Format Flag Use case
Terminal (default) Human review
JSON --format json IDE agents, dashboards, automation
YAML registry --format registry or --output-registry FILE toolcalls.yaml SBOM, PR diffs

Known limitations

  • Static analysis only — no runtime detection. If a guard is added by middleware or a gateway outside the file, annotate with // checked:ok — protected by [where].
  • Intra-procedural — guard detection looks at the same function or its immediate decorators. Cross-file guard chains require an annotation.
  • TypeScript files only.ts and .tsx files are scanned. Plain .js files are skipped silently. For Python agents, use diplomat-agent. The scanner emits a warning if no .ts files are found in the target directory.
  • ORM patterns require the import — Mongoose, Sequelize, and TypeORM use generic method names (.save(), .create()), so the patterns are scoped to files that import the ORM. Re-exported models may be missed.
  • Abstraction layers — if a repo wraps its ORM or HTTP client behind a custom module (e.g. db.ts re-exporting Prisma without a direct import 'prisma'), call sites in consumers won't carry the importContains scope and may be missed. Use // checked:ok at the wrapper boundary.
  • Large repos without tsconfig — scanning 5,000+ files without a tsconfig.json can take 30–60 s on slower machines (9 s on M-series for 7,874 files). Point at a subdirectory (scan ./src) to reduce scope.

Full limitations and pattern refinement history → docs/limitations.md

Roadmap

  • AST scanner with 40+ patterns across 12 categories
  • toolcalls.yaml behavioral SBOM with diff-stable output
  • OWASP Agentic Top 10 mapping
  • CI integration (--fail-on-unchecked)
  • // checked:ok annotations (with diplomat:ok / canary:ok aliases)
  • Validated against OpenClaw (7,874 files, ~9 s on M-series / ~30 s on x86, pinned commit 49d9996d)
  • Inter-procedural decorator resolution (v0.2)
  • SARIF 2.1.0 output (v0.2)
  • --diff-only for changed files (v0.2)
  • MCP server scanning (v0.3)
  • VS Code extension with inline diagnostics

Requirements

  • Node.js ≥ 20
  • 2 runtime dependencies: ts-morph (TypeScript compiler wrapper), yaml

Sibling projects

Community & support

Contributing

PRs welcome. The architecture above tells you exactly where things live. Read CONTRIBUTING.md first — it explains the "patterns are data, not logic" rule that keeps the matcher simple.

License

Apache 2.0 — Copyright 2026 Diplomat Services SAS. See LICENSE.