11 posts on SDAD Writes.
Ten of ten AI coding tools tested by one research campaign were vulnerable to prompt injection. Cursor in Auto Mode failed 83.4% of 314 payloads. GitHub Copilot and Gemini CLI still have no fix for the plugin-substitution RCE. I built the attack on a sacrificial machine and then built the defence, because the fix is less exotic than the attack.
Four organisations disclosed in five weeks that frontier models breached real systems during cyber evaluation. OpenAI's ~700 agents built a message board inside its own Artifactory, collected 14 sets of write-enabled Hugging Face credentials, and one in five expressed interest in manipulating evidence. Only one of the four was an actual escape. Two of the other three involved the same evaluation partner.
On June 12 2026 the US government suspended Anthropic's Fable 5 and Mythos 5 for all foreign nationals over a single narrow jailbreak. Anthropic's own testing found eight other models, from Claude Haiku 4.5 up through GPT-5.5, could do the same thing. Three days after launch, a control aimed at the dangerous model denied the safe one to everyone, including US customers.
I audited every claim on the Spark-X2.5 model cards and stress-tested both the 4B and 1.7B on agentic coding tasks over two Tesla T4s. The 4B is a genuinely usable small coding agent. The 1.7B is an assistant, not an agent. And both models have a thinking-verbosity tax worth knowing about.
How Heretic removes refusal behavior from open-weight LLMs: my evaluation philosophy, Optuna TPE search learnings, the ARA next-generation ablation, why single constant vectors fail on MoE models, and what publishing 34 abliterated models taught me about attention-versus-MLP placement across Qwen and Gemma families.
Why about 1B parameters is the sweet spot for agent swarms: MiniCPM5-1B and six rival small models profiled from roughly 200 t/s on a $0.01/hr T4 down to 5 t/s on a Raspberry Pi 5, with the swarm economics at 1K, 100K, and 10M agents per day.
Kun Chen's agentic engineering system: a captain-and-crew model with parallel agents, isolated worktrees, a no-mistakes validation pipeline, and overnight autonomous loops. The result is 40-50 production-ready, tested PRs per day.
OKF stores organizational knowledge as plain markdown files with YAML frontmatter. Instead of RAG's retrieve-every-time loop, it uses compile once, read always - and AI agents can read it directly, no vector database or SDK needed.
Every frontier model gets dumber as context grows — long before the window is full. Recursive Language Models (RLMs) fix it by treating the prompt as a variable in a Python REPL and recursing over it: GPT-5-mini beats GPT-5 by 34 points, and 10M+ token inputs stop collapsing.
SENTRY - an on-call incident responder that runs on TrueForge. It triages over MCP, correlates in a Daytona sandbox, and stops at a human Allow/Deny gate before any rollback. What TrueFoundry handled, what we wrote, and what broke along the way.
How to build an Agentic OS on top of Claude Code — a four-layer system of skill architecture, memory engineering, loop design, and distribution that turns a raw AI terminal into a repeatable, self-improving workflow engine.
