NVIDIA OpenShell and the case for agent controls outside the harness
NVIDIA's open-source OpenShell enforces agent policy outside the model and harness. What version 0.1.0 does, what is only a reference design, and what to copy.
5 min read
Blog tag
#Security
Threats, controls and secure design for AI tools and the systems they touch.
Using the OWASP Top 10 for Agentic Applications as an engineering checklist
How to turn the ten OWASP agentic AI risks, from goal hijack to rogue agents, into concrete backlog items for a coding-agent rollout this quarter.
6 min read
Protocol pivoting: when MCP trust leaks into another agent protocol
Research on protocol pivoting shows how an injected instruction can cross from MCP into agent-to-agent delegation. Why per-protocol fixes miss it.
SWE-Bench Pro Verified: when coding agents were reading the answers
SWE-Bench Pro Verified closes gold-patch leakage channels and refines 102 tasks. Some model scores fall sharply once agents can no longer fetch the solution.
4 min read
Claude Code hooks can now fail closed when enforcement breaks
Claude Code v2.1.295 adds onFailure: block for command and HTTP hooks so missing scripts, timeouts and unexpected exits deny the action instead of failing open.
MCP Python SDK redirected custom headers off-origin
A medium-severity flaw in the MCP Python SDK followed cross-origin redirects and forwarded custom API keys. How it differs from the TypeScript OAuth issuer bug.
What the Bank of England told firms about autonomous AI risk
The September 2026 FPC Record treats frontier AI autonomy and open-weight models as cyber and ops resilience issues for firms.
5 min read
Copilot local sandboxing is generally available: what it actually restricts
Copilot local sandboxing is GA in the CLI, app and VS Code Agent Host. What OS limits cover, leave out, and how they sit with policy.
4 min read
MCP TypeScript SDK OAuth flaw: credentials followed the wrong issuer
A high-severity MCP TypeScript SDK flaw let the server choose which authorization server got client secrets. Patch, bind and rotate.
Download counts aren't trust: a lesson for AI coding agents
A malicious copy of a popular maths library reached six million downloads in two days. Why agents that pick packages are exposed, and what limits the damage.
3 min read
What NIST is saying about AI agent security, and what to do before it lands
NIST summarised industry responses on AI agent security and is building control overlays for agents. The themes, the status, and steps to take now.
3 min read
Can another company's computer reach your server? Isolation, hop by hop
The seven checks a connection passes before it reaches a SpanDock server, which ones stop a stranger or another customer, and how to try it in the lab.
4 min read
Browse by tag
- #Security (55)
- #Agents (41)
- #AI coding tools (38)
- #LLM costs (23)
- #MCP (16)
- #Evaluation (14)
- #Observability (12)
- #Compliance (11)
- #Enterprise (9)
- #Platform engineering (9)
- #Data privacy (8)
- #DevOps (7)
- #Model routing (7)
- #Open-weight models (7)
- #Research (7)
- #Tutorials (6)
- #Code review (4)
- #Developer productivity (4)
- #Guardrails (4)
- #Networking (4)
- #OpenTelemetry (4)
- #Budgets (3)
- #Prompt engineering (3)
- #Self-hosting (3)
- #MCP gateway (2)
- #Release notes (2)
- #Skills (1)
- #Testing (1)