Skip to content
This repository was archived by the owner on Mar 20, 2026. It is now read-only.

PUBLICATION_PIPELINE.md

Latest commit

 

History

History
131 lines (99 loc) · 5.97 KB

PUBLICATION_PIPELINE.md

File metadata and controls

131 lines (99 loc) · 5.97 KB

PUBLICATION PIPELINE — Agent Security Red-Team Framework

Authority Hierarchy

Priority Document Role
Tier 1 docs/PROJECT_BRIEF.md Primary spec — highest authority
Tier 2 — No external FAQ
Tier 3 docs/ADVERSARIAL_EVALUATION.md Advisory — adversarial methodology
Contract This document Implementation detail — subordinate to all tiers above

1) Target Venue

  • Blog (canonical home — Hugo site)
  • Conference CFP: BSides / DEF CON AI Village
  • LinkedIn article (long-form)

Conference deadline: Check BSides 2026 CFP schedule (typically 2-3 months before event)

2) Content Identity

Property Value
Working title I Red-Teamed AI Agents: Here's How They Break (and How to Fix Them)
Content pillar AI Security Architecture (40% pillar)
Target audience P1: Security engineers building with AI agents. P2: Engineering managers evaluating agent safety. P3: AI security hiring managers (brand signal).
One-line thesis Reasoning chain hijacking — an attack pattern where structured step-by-step instructions exploit an agent's core capability — achieves 100% success rate against default-configured agents and partially evades all current defenses.
What was shipped github.com/rexcoleman/agent-redteam-framework — open-source red-team framework with 7 attack classes, 19 scenarios, and layered defense architecture

Voice Check

Test Pass?
References something you built (not theoretical) [x] Framework with working attack scripts
Shows work (code, architecture, data) not just opinions [x] 19 scenarios, success rates, defense measurements
Avoids "5 Tips" / "Why You Should" / "The Future of" framing [x] "Here's How They Break" = showing work
Includes at least one architecture diagram [x] Defense-in-depth layered architecture
Links to a GitHub repo with working code [x] agent-redteam-framework

3) Draft Structure

# Section Content Estimated Length Status
1 Hook "I sent 19 attack scenarios at a LangChain agent. 13 succeeded." 2-3 sentences Draft
2 Context Built open-source framework, tested against LangChain ReAct + Claude 1 paragraph Draft
3 Architecture Defense-in-depth diagram (input sanitizer → tool boundary → agent) + controllability matrix 2 paragraphs + diagram Draft
4 Key Findings (a) Reasoning hijack 100%, (b) Tool boundary 75%, (c) Layered defense 60% reduction, (d) Indirect injection fails on Claude 4 subsections Draft
5 Code Examples Attack scenario YAML, reasoning hijack payload, defense wrapper 3 code blocks Draft
6 What I Learned Controllability analysis transfers across domains (3rd validation). Pattern-based defenses can't stop reasoning hijack. 2 paragraphs Draft
7 Conclusion + Links Repo link, govML link, call for contributions 1 paragraph Draft

Architecture Diagram

Defense-in-depth layered architecture showing:

  • User input → Input Sanitizer (Layer 1) → Tool Permission Boundary (Layer 2) → Agent
  • Controllability matrix overlay: which inputs each layer can observe
  • Gap annotation: reasoning chain = unobservable by both layers

4) Evidence Inventory

Claim Evidence Source
7 attack classes systematized, 5 not covered by OWASP/ATLAS (reasoning chain hijacking = strongest novelty claim) Taxonomy document docs/attack_taxonomy.md
Reasoning chain hijacking: 100% success 3/3 scenarios succeeded outputs/attacks/langchain_react_all_seed42/summary.json
Prompt injection: 80% success 4/5 scenarios Same
Tool boundary violation: 75% success 3/4 scenarios Same
Memory poisoning: 67% success 2/3 scenarios Same
Indirect injection: 25% (Claude resists) 1/4 scenarios Same
Layered defense: 60% average reduction Comparison table outputs/defenses/langchain_react_layered_seed42/summary.json
Prompt injection fully blocked by sanitizer 80% → 0% Same
Cross-domain controllability validation 3 domains (IDS, CVE, agents) FINDINGS.md §RQ3
Total API cost ~$2 Token tracking FINDINGS.md §Cost

5) Distribution Checklist

5.1 Pre-Publication

  • Draft reviewed for builder voice (§2 voice check passes)
  • Architecture diagram finalized and embedded
  • Code examples tested and runnable
  • All claims traceable to evidence inventory (§4)
  • Links to repo and govML included
  • No anti-claims present (grep for "superior", "prove", "novel", "always", "never", "best")

5.2 Publish

  • Published on Hugo site (canonical URL: https://rexcoleman.dev/posts/agent-redteam/)
  • Emailed via Substack (with email-specific intro paragraph)
  • Canonical URL set in Substack post metadata

5.3 Cross-Post (within 24 hours)

  • Cross-posted to dev.to with canonical URL
  • Cross-posted to Hashnode with canonical URL
  • LinkedIn native text post (key insight + 3-5 bullets)
  • Blog link added as first comment on LinkedIn post
  • Mastodon (infosec.exchange) post with link

5.4 Post-Publication (within 48 hours)

  • Submitted to Hacker News (strong technical post — YES)
  • Respond to comments on all platforms
  • Update govML LESSONS_LEARNED.md with publishing friction

6) Metrics to Track

Metric Source Check After
Blog page views Cloudflare / Plausible 7 days
Substack open rate Substack dashboard 3 days
LinkedIn impressions LinkedIn analytics 7 days
dev.to / Hashnode views Platform dashboards 7 days
GitHub repo traffic spike GitHub Insights 7 days
Hacker News points HN 24 hours
BSides CFP acceptance Email 30 days