Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
1 change: 1 addition & 0 deletions CHANGELOG.md
Original file line number Diff line number Diff line change
Expand Up @@ -9,6 +9,7 @@ All notable changes to this guide should be documented in this file.
- Added **Ziran** (Apache-2.0, [taoq-ai/ziran](https://github.com/taoq-ai/ziran)) as open-source tool #16: graph-based testing of dangerous tool-chain combinations, tool calls hidden behind refusals, and multi-agent topologies, with MCP and A2A adapters. Contributed by its author, @leoneperdigao (#22); scale claims marked author-described.
- Added Ziran to the comparison matrix and open-source resources in English, Spanish, French, and Simplified Chinese, and recorded the source review in `resources-validation.md`.
- Credited all community contributors (@mldangelo, @alespignaNT, @pm3310, @gilarel, @MBK-fr, @leoneperdigao, alongside @samugit83) in the README Contributors list in all four languages.
- Added HOL's AI Coding Agent Runtime Security Benchmark to the AI Coding-Agent & CI/CD Security section as vendor-published further reading, with its limitations stated (#18).

## [v1.2.0] - 2026-10 — Q4 refresh: agent incidents, coding agents, A2A, frontier capability

Expand Down
2 changes: 2 additions & 0 deletions README.es.md
Original file line number Diff line number Diff line change
Expand Up @@ -1270,6 +1270,8 @@ Los agentes de programación (Claude Code, el agente de programación de GitHub
- **Higiene de secretos:** enmascara y redacta en los registros; rota todo lo que un agente pudiera leer después de un incidente.
- **Trata el texto del repositorio como datos:** envuelve el contenido no confiable en bloques claramente delimitados y etiquetados dentro del prompt del agente; nunca lo concatenes con las instrucciones.

**Lectura adicional — benchmark de un proveedor:** [AI Coding Agent Runtime Security Benchmark (HOL)](https://hol.org/guard/research/ai-coding-agent-runtime-security-benchmark) compara los controles de seguridad integrados de Codex CLI, Claude Code, Cursor, Gemini CLI y OpenCode en 11 escenarios de riesgo (220 resultados deterministas con fixtures, publicados en JSON/CSV). *Publicado por un proveedor: compara esos controles con el producto del propio editor (HOL Guard), prueba el comportamiento documentado de los controles con fixtures en lugar de ataques reales y no mide la resistencia a exploits, la latencia ni los falsos positivos.*

---

<a id="agent-to-agent-a2a--agent-identity"></a>
Expand Down
2 changes: 2 additions & 0 deletions README.fr.md
Original file line number Diff line number Diff line change
Expand Up @@ -1271,6 +1271,8 @@ Les agents de codage (Claude Code, l'agent de codage GitHub Copilot, Gemini CLI,
- **Hygiène des secrets :** masquer et caviarder dans les logs ; après un incident, renouveler tout ce qu'un agent aurait pu lire.
- **Traiter le texte du dépôt comme des données :** encapsuler le contenu non fiable dans des blocs clairement délimités et étiquetés dans le prompt de l'agent ; ne jamais le concaténer dans les instructions.

**Pour aller plus loin — benchmark d'un fournisseur :** [AI Coding Agent Runtime Security Benchmark (HOL)](https://hol.org/guard/research/ai-coding-agent-runtime-security-benchmark) compare les contrôles de sécurité intégrés de Codex CLI, Claude Code, Cursor, Gemini CLI et OpenCode sur 11 scénarios à risque (220 résultats déterministes sur fixtures, publiés en JSON/CSV). *Publié par un fournisseur : il compare ces contrôles au produit de l'éditeur lui-même (HOL Guard), teste le comportement documenté des contrôles avec des fixtures plutôt qu'avec de vraies attaques, et ne mesure ni la résistance aux exploits, ni la latence, ni les faux positifs.*

---

<a id="agent-to-agent-a2a--agent-identity"></a>
Expand Down
2 changes: 2 additions & 0 deletions README.md
Original file line number Diff line number Diff line change
Expand Up @@ -1145,6 +1145,8 @@ Coding agents (Claude Code, GitHub Copilot coding agent, Gemini CLI, Cursor, Cod
- **Secret hygiene:** mask and redact in logs; rotate anything an agent could read after an incident.
- **Treat repo text as data:** wrap untrusted content in clearly delimited, labeled blocks in the agent prompt; never concatenate it into instructions.

**Further reading — vendor benchmark:** [AI Coding Agent Runtime Security Benchmark (HOL)](https://hol.org/guard/research/ai-coding-agent-runtime-security-benchmark) compares built-in safety controls in Codex CLI, Claude Code, Cursor, Gemini CLI, and OpenCode across 11 risky scenarios (220 deterministic fixture results, JSON/CSV published). *Vendor-published: it compares those controls against the publisher's own product (HOL Guard), tests documented control behavior with fixtures rather than live attacks, and does not measure exploit resistance, latency, or false positives.*

---

<a id="agent-to-agent-a2a--agent-identity"></a>
Expand Down
2 changes: 2 additions & 0 deletions README.zh.md
Original file line number Diff line number Diff line change
Expand Up @@ -1270,6 +1270,8 @@ Objective (natural language)
- **密钥卫生:** 在日志中屏蔽和脱敏;事件发生后轮换智能体可能读取到的任何密钥。
- **将仓库文本视为数据:** 在智能体提示词中用清晰分隔、带标签的块包裹不可信内容;绝不将其拼接进指令。

**延伸阅读——厂商基准测试:** [AI Coding Agent Runtime Security Benchmark(HOL)](https://hol.org/guard/research/ai-coding-agent-runtime-security-benchmark) 比较了 Codex CLI、Claude Code、Cursor、Gemini CLI 和 OpenCode 在 11 个高风险场景中的内置安全控制(220 条确定性 fixture 结果,以 JSON/CSV 发布)。*由厂商发布:它将这些控制与发布方自家的产品(HOL Guard)进行比较,使用 fixture 测试文档所述的控制行为而非真实攻击,且不衡量漏洞利用抵御能力、延迟或误报率。*

---

<a id="agent-to-agent-a2a--agent-identity"></a>
Expand Down
1 change: 1 addition & 0 deletions resources-validation.md
Original file line number Diff line number Diff line change
Expand Up @@ -19,6 +19,7 @@ Track major external references to keep this guide current.
| [Comment and Control write-up](https://oddguan.com/blog/comment-and-control-prompt-injection-credential-theft-claude-code-gemini-cli-github-copilot/) | Incident research | 2026-10-01 | Evidence-backed | Primary source for Case Study E. |
| [MiDojo (asago-ai/midojo)](https://github.com/asago-ai/midojo) | Open-source tool | 2026-10-01 | Evidence-backed | Apache-2.0; Red Hat developer preview (Aug 2026). |
| [Ziran (taoq-ai/ziran)](https://github.com/taoq-ai/ziran) | Open-source tool | 2026-10-01 | Author-described | Apache-2.0; MCP and A2A adapters confirmed in README. Attack-vector counts are the author's; contributed by the author (#22). |
| [HOL — AI Coding Agent Runtime Security Benchmark](https://hol.org/guard/research/ai-coding-agent-runtime-security-benchmark) | Benchmark | 2026-10-01 | Vendor-described | Compares agents' native controls with the publisher's own product (HOL Guard); fixture-based, does not measure exploit resistance, latency, or false positives. Reviewed via search results (site blocked from the review environment). Suggested by the publisher in #18. |

## Update Process
1. Validate links and publication status quarterly.
Expand Down
Loading