Skip to content
Roadmap0/111
not startedstudy day 3 of 6 · covers original day 42

Add audit events, timeouts, cancellation, and error boundaries…

P3 Agent systemS3D2
not started0/5 blocks
0/5
Jump & context

Week 6: MCP Proxy Hardening, LangGraph Checkpoints, OAuth2 Delegation & Inter-Agent Bus

The five hours

  1. 01

    Concepts & system design

    60 min

    R21 — MCP cancellation & error handling

    R27 — OpenTelemetry context propagation & audit logs

    Read the assigned sections; map the mechanism, threat model and trade-offs.

  2. 02

    Project implementation

    120 min

    Add structured audit event emission, per-call timeouts, output-size caps, cancellation propagation, and egress restrictions to D2.

    Build, integrate, debug and improve the increment in the active project.

  3. 03

    Security lab

    60 min

    Run combined prompt-injection, malformed JSON-RPC, timeout, and unauthorized tool tests through the agent D2 MCP server path.

    Test the abuse case, failure path, authorization bypass or resource limit.

  4. 04

    Verification & evidence

    30 min

    Integrated D2 Sandboxed MCP Proxy security test report.

    Add or run regression tests; record the 9-point evidence record.

  5. 05

    Deliberate practice & Socratic review

    30 min

    Answer the 8 Socratic questions without notes: What? Why? How? What if? Why not? How can I prove it? Where can it break? What have I learned?

    Re-explain the hardest concept aloud, inspect one edge case, update the gap log.

  6. total5 h

Read today · inside the concepts block

R21 R27

Read only the named sections needed for today's task. Reading sits inside the 60-minute concepts block; it is not extra time.

Free resources for this day

5 curated
  • docsOpenTelemetry · ~10 min read
    Context propagation

    Explains trace/span context injection and extraction across service boundaries and the security warning about forged trace headers and leaking baggage — the correlation spine for audit events.

    R27moved 2026-11-11
  • videoLightstep is now ServiceNow Cloud Observability · ~10 min
    Context Propagation makes OpenTelemetry awesome

    Walks why traces break without propagation and how the context object flows through nested spans, so per-call spans for each MCP invocation stay in one trace.

    R27checked 2026-11-11
  • videoOpenTelemetry · ~38 min
    OTEL in Practice: Context Propapagtion in Distributed Tracing with Doug Ramirez

    Python-specific walkthrough of inject/extract and manual instrumentation, including where to attach attributes to spans rather than dumping payloads into them.

    R27checked 2026-11-11
  • articleOWASP · ~15 min read
    MCP Security - OWASP Cheat Sheet Series

    Prescribes logging and centrally monitoring every tool invocation, plus output-size and least-privilege controls, which are the audit-event and egress requirements for the proxy.

    R21checked 2026-11-11
  • labOpenTelemetry Python · 3 runnable scripts
    Basic Context — OpenTelemetry Python documentation

    Runnable implicit, child and async context examples let you prove spans nest correctly around each tool call, including the cancellation path in an async handler.

    R27checked 2026-11-11

Evidence record · 9-point standard

0/9

A lab is complete only when it produces evidence. Installing PyRIT, Garak, Presidio or TransformerLens is not completion. Saved on this device only.

Day notes

empty

The two source formulations

verbatim · nothing summarised

Project & Deliverable Target

P3 | `D2` (Sandboxed MCP Proxy) | S3

Core Implementation

Add structured audit event emission, per-call timeouts, output-size caps, cancellation propagation, and egress restrictions to D2.

Security Lab

Run combined prompt-injection, malformed JSON-RPC, timeout, and unauthorized tool tests through the agent D2 MCP server path.

Verification & Evidence

Integrated D2 Sandboxed MCP Proxy security test report.

Read Today (60m Concepts Block)

R21 (MCP cancellation & error handling), R27 (OpenTelemetry context propagation & audit logs).

---