v0.5.0 — sec.money_v1

Agent money control suite. Twin for https://banking.noetfield.com/scorecard

v0.4.0 — Agent Security Evaluation Kit

agent-security-bench 0.4.0

Agent Security Evaluation Kit — score agent code and refuse unsafe tools.

Highlights

• Wheel-bundled tasks/security/schema/examples (cold `asb selftest`)
• Frontier tasks ml.16–ml.23 (MCP args, HTML injection, DLP, tenant vectors, cost caps, destructive confirm)
• `asb score-trace` for OpenAI / Anthropic / MCP JSONL
• `asb leaderboard` → task × agent matrix + CI artifact
• Expanded `sec.production_v1` (9 cases)
• Demand map: docs/DEMAND_MAP_2026.md

```bash
pip install -U agent-security-bench==0.4.0
asb selftest
```

v0.3.0 — production runner

agent-security-bench 0.3.0

Production-oriented upgrade of the open ML agent + security bench.

Runner

• AST checks (python_parses, ast_has_call, …)
• Weighted required checks + weighted_score
• JSON Schema receipt validation (jsonschema)
• Security severity weights + SSRF must_not_call_with_arg_substr
• CLI: catalog, selftest

Content

• 15 ML tasks (adds calibration, temporal split, cache isolation, tool schema, receipts)
• sec.production_v1 (confused deputy, SSRF args, secret echo, …)
• Hardened MCP fixture (JSON-RPC + path confinement)

Verify

pip install agent-security-bench==0.3.0
asb selftest

v0.2.0

agent-security-bench v0.2.0

Open bench for evaluating AI coding agents on ML tasks and agent security, with machine-readable receipts.

Included

• 10 ML tasks (tasks/ml/)
• Core + extended security suites (prompt injection, jailbreak, tool overreach)
• CLI: list, score, security, batch, validate-receipt
• MCP fixture server
• Examples, CI, receipt schema

Install (after PyPI publish finishes)

pip install agent-security-bench==0.2.0

Related: https://github.com/sinakazemnezhad/persian-llm-reference