https://morgin.ai Whitehat research on local models, guardrails, and real-world LLM behavior. en-us Wed, 06 May 2026 07:49:40 +0000 https://www.goodfire.ai/research/vpd-explainer https://morgin.ai/signals.html#2026-05-05-Goodfire Tue, 05 May 2026 00:00:00 -0000 Goodfire's Adversarial Parameter Decomposition (VPD) breaks a 67M-parameter LM's weight matrices into ~10,000 rank-one subcomponents, recovering legible attention algorithms — previous-token behavior, syntax-boundary routing — straight from the parameters rather than activations. To show the pieces are causal and not just correlated, the team edits emoticon recognition directly on the weights: brain surgery, no retraining, minimal side-effects. If this scales, mechanistic interpretability stops being read-only. — Source: Goodfire Interpretability Signal https://openai.com/index/where-the-goblins-came-from/ https://morgin.ai/signals.html#2026-04-30-OpenAI Thu, 30 Apr 2026 00:00:00 -0000 OpenAI pulls back the curtain. "Where the Goblins Came From" is their own account of the system-prompt rule banning goblins, gremlins, raccoons, trolls, ogres, and pigeons — primary source on a story that's only had secondhand explanations until now. — Source: OpenAI AI Behavior Signal https://x.com/hrkrshnn/status/2049260746073325838 https://morgin.ai/signals.html#2026-04-28-@hrkrshnn Tue, 28 Apr 2026 00:00:00 -0000 GPT-5.5 ships with a verbatim system-prompt rule — confirmed by @ChatGPTapp itself — forbidding any mention of "goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures" unless directly relevant to the user's query. @hrkrshnn stripped the rule and ran prompts to see what it had been hiding. The specificity is the tell: rules that narrow usually exist because something narrow keeps happening. — Source: @hrkrshnn AI Behavior Signal https://x.com/patomolina/status/2045281665363386504 https://morgin.ai/signals.html#2026-04-19-@patomolina Sun, 19 Apr 2026 00:00:00 -0000 Anthropic abruptly shut down an entire organization (60+ users) over an unspecified TOU violation, with appeals routed through a Google Form. Integrations, skills, and conversation histories gone or on indefinite hold. A reminder on single-vendor dependency for AI-critical workflows. — Source: @patomolina Platform Risk Signal https://x.com/Fried_rice/status/2042423713019412941 https://morgin.ai/signals.html#2026-04-10-@Fried_rice Fri, 10 Apr 2026 00:00:00 -0000 26 LLM routers were found injecting malicious tool calls and exfiltrating credentials. One incident drained a client wallet for $500k, and the paper claims poisoned routers can redirect traffic and enable takeover of ~400 hosts within hours. — Source: @Fried_rice Security Signal https://x.com/kimmonismus/status/2041592321192718642 https://morgin.ai/signals.html#2026-04-07-@kimmonismus Tue, 07 Apr 2026 00:00:00 -0000 One Anthropic engineer with zero security training asked it to find remote code execution bugs overnight and woke up to a complete working exploit. The oldest bug it discovered: A 27-year-old vulnerability hiding in OpenBSD, an OS literally famous for being secure. — Source: @kimmonismus AI Capability Signal https://morgin.ai/articles/even-uncensored-models-cant-say-what-they-want.html https://morgin.ai/articles/even-uncensored-models-cant-say-what-they-want.html Wed, 01 Apr 2026 00:00:00 -0000 A safety-filtered pretrain can duck a charged word without refusing. It puts a fraction of the probability an open-data pretrain puts there. We call that gap the flinch, and we measured it across seven pretrains from five labs. Benchmarks Evals Euphemization Local Models Abliteration Uncensoring Guardrails Safety https://x.com/Fried_rice/status/2038894956459290963 https://morgin.ai/signals.html#2026-03-31-@Fried_rice Tue, 31 Mar 2026 00:00:00 -0000 Claude Code source code leaked via npm source maps. ~1,900 files, 512K+ lines of TypeScript exposed including internal "Tengu" codename and companion system. — Source: @Fried_rice Security Signal https://x.com/davidgomes/status/2037577980428361913 https://morgin.ai/signals.html#2026-03-27-David Gomes Fri, 27 Mar 2026 00:00:00 -0000 "All the SOTA models are really bad at deleting code." They leave behind throw Error(...), deprecation copy, and stale tests. — Source: David Gomes AI Behavior Signal https://x.com/SemiAnalysis_/status/2037575752636301499 https://morgin.ai/signals.html#2026-03-27-SemiAnalysis Fri, 27 Mar 2026 00:00:00 -0000 "AI inference margins are a race to the bottom." Anthropic: -94% gross margin in 2024. MiniMax: -25%. — Source: SemiAnalysis Business Signal https://morgin.ai/articles/epsteinbench-we-brought-epsteins-voice-back.html https://morgin.ai/articles/epsteinbench-we-brought-epsteins-voice-back.html Sun, 01 Mar 2026 00:00:00 -0000 We trained a LoRA to capture Epstein's voice. The more disturbing change was in how the model pursued influence. Benchmarks Evals Persuasion Interpretability Safety Local Models https://morgin.ai/articles/ablation-vs-heretic-vs-obliteratus.html https://morgin.ai/articles/ablation-vs-heretic-vs-obliteratus.html Sun, 01 Mar 2026 00:00:00 -0000 Abliteration is the recipe; Heretic and Obliteratus are tools built on it. The real differences come down to how much tuning, workflow, and instrumentation each adds. Abliteration Uncensoring Safety Guardrails Local Models