Articles
Long-form writing on tech, culture, and the edges of the internet.
The benchmark score is the number to trust least
How to weigh Claude Opus 5.5's intelligence, latency, and token cost, and where its real AI safety and cybersecurity risks concentrate.
the phone rings in a voice you trust
Unreal Agent clones trusted voices for social engineering, and any identity check that ends at voice recognition is already defeated.
Your telemetry toggle was a file access control
Claude Code read AGENTS.md only when telemetry was on. Binding access to a diagnostic flag breaks least privilege. The fix closed the instance, not the class.
An unsolved cipher was never a secure one
GPT-6 Astra solved an Enigma message unsolved since 2005. The failure was effort-based security, not the cipher, and that pattern reaches legacy encryption.
Every Prompt Is A Retrieval Request
Meta's Muse returned 6.8GB when asked for its filesystem. The response channel serviced a request for internal data with no enforcement at the boundary.
Grok 4.7 collapses recon into a scripted session
Grok 4.7's real-time retrieval and agentic tool-use compress red team reconnaissance and phishing, degrading content-based detection while behavior-based rules hold.
Heretic strips refusals from open-weight models
Heretic automates stripping refusals from open-weight LLMs. Why model-level guardrails were never a security control, and what defenders should do instead.
Mini-AGI trains a model you can't sign
Mini-AGI trains a continual-learning model on 8GB VRAM. Why self-updating weights break signing, auditing, and red-team results, and how to defend.
One ../ and the server hands over the wrong file
Unauthenticated WordPress path traversal lets an anonymous request select which file the server reads; execution follows where a code-executing sink exists.
Your watermark is a spymark
Watermark detection recovers a signal, not an author, so any party with write access can embed a hidden tracking mark defender tooling reads as clean.
A privacy toggle is a sign, not a lock
A security analysis of Grok 4.7's privacy versus other AI assistants, judged by enforcement point rather than feature count.
Cloudflare ships Python Workers to general availability
How to wire Cloudflare Python Workers into automation pipelines as validated, Queue-connected stages for reliable, scalable AI systems.