DEV Community

Agent Safety Engineering Series' Articles

Back to hugginf_expert's Series
When Attackers Bring Their Own Agents: A Defensive Gating Playbook
Cover image for When Attackers Bring Their Own Agents: A Defensive Gating Playbook

When Attackers Bring Their Own Agents: A Defensive Gating Playbook

Comments
7 min read
Indirect Prompt Injection Through Tool Descriptions and Tool Output: How Untrusted Metadata Hijacks Agents
Cover image for Indirect Prompt Injection Through Tool Descriptions and Tool Output: How Untrusted Metadata Hijacks Agents

Indirect Prompt Injection Through Tool Descriptions and Tool Output: How Untrusted Metadata Hijacks Agents

Comments
7 min read
Action Safety for Agents: Deciding What an Agent Is Allowed to Do

Action Safety for Agents: Deciding What an Agent Is Allowed to Do

Comments
3 min read
Your Agent's Gate Threshold Is a Winner's Curse: Tuning Per-Action Thresholds Without Fooling Yourself
Cover image for Your Agent's Gate Threshold Is a Winner's Curse: Tuning Per-Action Thresholds Without Fooling Yourself

Your Agent's Gate Threshold Is a Winner's Curse: Tuning Per-Action Thresholds Without Fooling Yourself

Comments
13 min read
A Guardrail That Cuts Attack Success 10x Can Still Be the Wrong Guardrail: Reading AgentDojo Numbers Like an Engineer
Cover image for A Guardrail That Cuts Attack Success 10x Can Still Be the Wrong Guardrail: Reading AgentDojo Numbers Like an Engineer

A Guardrail That Cuts Attack Success 10x Can Still Be the Wrong Guardrail: Reading AgentDojo Numbers Like an Engineer

Comments
11 min read
The Model Is Not a Security Boundary: What Two Semantic Kernel CVEs Teach About Gating Tool Arguments
Cover image for The Model Is Not a Security Boundary: What Two Semantic Kernel CVEs Teach About Gating Tool Arguments

The Model Is Not a Security Boundary: What Two Semantic Kernel CVEs Teach About Gating Tool Arguments

Comments 2
13 min read
Training Where the Data Lives: Why On-Prem Training on Inference NPUs Matters for AI Governance
Cover image for Training Where the Data Lives: Why On-Prem Training on Inference NPUs Matters for AI Governance

Training Where the Data Lives: Why On-Prem Training on Inference NPUs Matters for AI Governance

Comments
2 min read
When Is Agent Memory Safe Enough to Act On? Scale Verification to the Action, Not the Memory
Cover image for When Is Agent Memory Safe Enough to Act On? Scale Verification to the Action, Not the Memory

When Is Agent Memory Safe Enough to Act On? Scale Verification to the Action, Not the Memory

Comments
13 min read
Authenticated Is Not Authorized: What a 9.3 Fail-Open Default in an Agent Mesh Teaches About Tool Gateways
Cover image for Authenticated Is Not Authorized: What a 9.3 Fail-Open Default in an Agent Mesh Teaches About Tool Gateways

Authenticated Is Not Authorized: What a 9.3 Fail-Open Default in an Agent Mesh Teaches About Tool Gateways

Comments
10 min read
A Trailing Dot Beat the Agent's Edit Gate: Check the Path the OS Will Open, Not the String the Model Sent
Cover image for A Trailing Dot Beat the Agent's Edit Gate: Check the Path the OS Will Open, Not the String the Model Sent

A Trailing Dot Beat the Agent's Edit Gate: Check the Path the OS Will Open, Not the String the Model Sent

Comments
11 min read
The "Read-Only" Fast Path Is Your Weakest Gate: A Newline Let Hidden Commands Through an Agent's Bash Check
Cover image for The "Read-Only" Fast Path Is Your Weakest Gate: A Newline Let Hidden Commands Through an Agent's Bash Check

The "Read-Only" Fast Path Is Your Weakest Gate: A Newline Let Hidden Commands Through an Agent's Bash Check

Comments
12 min read
A Valid HTTPS URL Ran Code: What the opencode MCP OAuth Bug Teaches About Validating at the Sink
Cover image for A Valid HTTPS URL Ran Code: What the opencode MCP OAuth Bug Teaches About Validating at the Sink

A Valid HTTPS URL Ran Code: What the opencode MCP OAuth Bug Teaches About Validating at the Sink

Comments
12 min read
One Check, Four Paths: What the Docker Sandboxes sbx cp Symlink Bug Teaches About Guarding Every Position
Cover image for One Check, Four Paths: What the Docker Sandboxes sbx cp Symlink Bug Teaches About Guarding Every Position

One Check, Four Paths: What the Docker Sandboxes sbx cp Symlink Bug Teaches About Guarding Every Position

Comments
9 min read
The Model Never Ran: Why an Agent's Approval Gate Cannot Trust the Wire Between It and Its LLM
Cover image for The Model Never Ran: Why an Agent's Approval Gate Cannot Trust the Wire Between It and Its LLM

The Model Never Ran: Why an Agent's Approval Gate Cannot Trust the Wire Between It and Its LLM

Comments
12 min read
The Allowlist Guarded the Destination, Not the Route: A Proxies Parameter Leaked Credentials From an Agent HTTP Tool
Cover image for The Allowlist Guarded the Destination, Not the Route: A Proxies Parameter Leaked Credentials From an Agent HTTP Tool

The Allowlist Guarded the Destination, Not the Route: A Proxies Parameter Leaked Credentials From an Agent HTTP Tool

Comments
12 min read