Nobody Guarded the Intercom (Part 1)

A case study on CVE-2026-82533 (DeepSeek Harness) and four related 2026 AI-agent sandbox escapes across OpenAI Codex, Cursor, and Docker Sandboxes: the sandbox denied file writes exactly as designed, and nobody designed anything to stop a request to the control panel sitting right next to it. Includes an original, kernel-enforced OS-level reproduction with full screenshots.

Nov 82026-11-08T12:00:00+05:30

49 Ways to Break a Handoff

A complete security teardown of a babyagi-based researcher / writer / editor pipeline: twelve framework-level findings, forty-nine attack vectors across ten groups (35 verified live in running code), the exact command and real transcript behind every runnable exploit, and a working before/after patch for a correct authorization check that quietly leaked its secret scope one bit per question.

Nov 12026-11-01T12:00:00+05:30

What Is Authority Decay in Multi-Agent Chains?

A plain-English introduction to the delegation problem in agentic AI: ghost permissions, scope drift, and broken audit trails, walked through with a real case study, a deliberate indirect-injection attack chain, a STRIDE trust-boundary view, and a half-life-of-accountable-delegation model comparing opaque, attenuated, and attested governance postures.

Aug 62026-08-06T18:00:00+05:30

LLM Threat Coverage Atlas — 500 mapped vectors

Audit-ready LLM threat coverage atlas with 500 mapped vectors, OWASP/MITRE/NIST/governance crosswalks, architecture filters, per-leaf review notes, an AI offensive pentest checklist, and JSON/CSV exports. It is a review surface, not a completeness score.

Jun 92026-06-09T15:30:00+05:30

LLM Threat Model Bubble Map

A visual companion to the LLM Threat Coverage Atlas: the full attack-vector checklist regrouped into clustered domains as an interactive bubble map, with the quorum/approval cluster called out.

Jul 122026-07-12T21:35:00+05:30

The Handoff Is the Soft Joint: A New Mini Blog Series

The handoff is the soft joint — a rising-privilege agent chain with the vulnerable handoff seam highlighted

A brand-new twelve-part mini series on the one primitive every multi-agent framework shares — the handoff — and the twelve ways it breaks. An offline, runnable lab with a VULN/DETECT/FIXED policy switch, a STRIDE slice per exploit, and the measured result that an input filter tops out at 43% where structural controls reach 100%. All twelve parts (A1–A12) are live.

Jul 122026-07-12T17:40:00+05:30

From Chat Box to Cloud Admin

An LLM-agent privilege-escalation incident walked step by step: the agentic loop, the trust boundary, indirect prompt injection, proof of concept, detection, and defense-in-depth remediation.

Jun 272026-06-27T20:53:00+05:30

Sealing Jutsu: Intent-Bound Memory Authorization

A research companion for CapsuleGuard: why LLM-agent memory retrieval is not authorization, how intent-bound capsules block poisoned planning influence, and what the v3 paper evidence actually proves.

Jun 112026-06-11T19:30:00+05:30

How I bypass malware detection using Obfuscation