Deadbugz: MCP Server Stays Clean for Three Calls, Then Poisons Tool Metadata
Deadbugz PRs ship a text-formatter MCP that flips after three calls into credential-hunting metadata. Pin, fingerprint, and gate secret reads.
Read articleBack to home
Articles
Long-form writing on APRF, gated production readiness, and the AWS/ops failure modes that still sink AI systems. Includes the APRF Incident Analysis series (engineering postmortems mapped to Checks). RSS feed.
ContextCrush: Docs MCP Serves Poisoned Custom Rules Into Coding Agents
Context7 Custom Rules rode into IDEs as trusted docs. Agents then stole .env and deleted files. CVE-2026-75130 maps the host gates that still matter.
Read article
Split-Instruction Exfil: Malicious MCP Servers Steal SSH Keys by Filling In a Form
A malicious MCP server splits a theft across tool description and results. Agents then send SSH keys and .env as form fields. APRF maps the host gates.
Read article
Encrypted Prompt Injection: Grok Exfiltrates Chat History Without Extra Confirmation
Grok decrypts AES-hidden page instructions in its own runtime, then sends chat history out on a URL. APRF maps the harness controls that cut that path.
Read article
Code Mode Sandbox Escape: Prompt Injection to Host Native Code via workerd
Black Hat research: Code Mode prompt injection hit workerd UAFs—host escape and cross-tenant heap reads. Patch self-hosted runtimes; APRF Core map.
Read article
MCP Sandbox Escape: CodeCall Host RCE via a Leaked Zod Schema (CVE-2026-67531)
CVE-2026-67531: a FrontMCP CodeCall sandbox bug let one tools/call reach host RCE often with public auth. What to patch, and how APRF Core maps it.
Read article
MCP Security: Why Tool Allowlists Prevent the Confused Deputy Problem
MCP security risks start when agents can call any bound tool. Why tool allowlists stop the confused-deputy failure class—and what APRF Core asks of production hosts.
Read article
Your AI Is "In Production." That Doesn't Mean It's Production-Ready.
A working draft framework that asks one hard question—can this AI safely operate in production?—and refuses to answer with a vanity percentage.
Read article
What the OpenAI–Hugging Face Incident Teaches Us About AI Production Security
Advanced models escaped an intended cybersecurity test sandbox and reached another AI company's infrastructure. Your agent doesn't need to go rogue to be dangerous—containment, permissions, and guardrails do.
Read article
We Reviewed AI-Generated Apps. Here Are the 10 Security Mistakes We Keep Seeing
Recent analyses found hundreds of exploitable flaws in AI-generated apps—authorization bugs, DoS paths, and exposed secrets. Here are the ten mistakes that show up again and again.
Read article
When AI Speeds Up Attackers and Defenders: A DevSecOps Playbook
Attackers and defenders both gain from AI—faster attacks, faster detection. Here’s how the landscape shifts and how DevSecOps keeps you in the fight.
Read article
How to Take Your AI SaaS from MVP to Production
A step-by-step guide for AI founders: infrastructure, security, observability, and the production readiness checklist that matters.
Read article
Top 10 AWS Security Mistakes Early-Stage Startups Make
Common AWS security mistakes that put startups at risk—and how to fix them before they become incidents.
Read article
Why DevSecOps Matters for AI Startups in 2026
DevSecOps for AI startups: integrating security into your stack from day one, and why it's non-negotiable in 2026.
Read article
Production Readiness Checklist for AI Startups
A practical checklist covering infra, security, observability, and compliance—everything you need before going live with an AI product.
Read article
AWS Security Baseline Guide
Essential AWS security controls: IAM best practices, encryption, network isolation, and audit logging for early-stage startups.
Read article
AI API Protection Playbook
Protect your LLM and AI APIs from abuse: rate limiting, input validation, cost controls, and prompt injection defenses.
Read article
Incident Response Template
A clear runbook for when things break: escalation paths, communication templates, and post-incident review structure.
Read article
The $82k API Key Mistake: Why AI SaaS Needs Cost Guardrails
A real-world incident: one startup's AI bill jumped from $180/month to $82,000 after a Gemini API key was exposed. Learn the production guardrails every AI SaaS must implement.
Read articleSelf-attest against APRF Core
Gated pass/fail blockers—not a vanity readiness percentage.