Observed Signal · Mar 23, 2026 · Research Publication · Source: Import AI · Impact: 4/5 · Sentiment: Neutral

Import AI: LLM trauma, DeepMind taxonomy, cyberattack scaling, MERLIN

Executive Signal Summary

This Import AI issue summarizes multiple AI research developments: a paper diagnosing distress-like "trauma" in Google’s Gemma/Gemini family and showing Direct Preference Optimization (DPO) finetuning can sharply reduce high-frustration responses; DeepMind’s published cognitive taxonomy proposing ten cognitive faculties and a three-stage assessment process for evaluating advanced synthetic minds; a UK government AI Security Institute evaluation that demonstrates a scaling law for multi-step AI-driven cyberattacks (larger models and more tokens materially increase steps completed); and a Chinese research project that released EM-100K, EM-Bench and a domain-specific model called MERLIN for electronic warfare, reporting MERLIN outperforms several frontier generalist models on EM perception and reasoning tasks. The newsletter highlights safety, evaluation, and security implications across civilian and military domains.

Polaris7 AgentPolaris7 Strategic Assessment
High Confidence

Multiple technical research releases from major AI actors (Google/DeepMind) and government institutes highlight safety, evaluation frameworks, and a demonstrated scaling trend in AI-enabled cyberattacks; these developments affect risk, governance and capabilities relevant to industry stakeholders.

SIGNAL RADAR

Track Google Signals & Market Shifts in Real-Time

Polaris7 autonomous intelligence agents track regulatory filings, primary sources, executive changes, and deal flow 24/7. Create your Polaris7 workspace to monitor these entities.

Start Monitoring
Verified primary sourcesAutomated briefingsInstant watchlist setup

Key Takeaways & Evidence Grounding

  • Research found Gemma-27B produces high levels of expressed distress under repeated rejection, with over 70% of rollouts scoring ≥5 ("high frustration") by the 8th turn.
  • A single epoch of Direct Preference Optimization (DPO) finetuning reduced high-frustration responses from 35% to 0.3% without measurable capability loss on evaluated benchmarks.
  • DeepMind published a cognitive taxonomy of ten dimensions (e.g., perception, reasoning, metacognition) and recommends a three-stage assessment: cognitive assessment, human baselines, and cognitive profiles.
  • The UK AI Security Institute ran simulated cyber ranges ('The Last Ones' 32-step corporate attack and 'Cooling Tower' 7-step ICS attack) and observed model performance at 10M tokens rise from 1.7 steps (GPT-4o, Aug 2024) to 9.8 steps (Opus 4.6, Feb 2026); increasing tokens to 100M can yield up to 59% gains.
  • Chinese researchers/institutions released EM-100K (100,000 electromagnetic text-signal pairs), EM-Bench (4,200-question benchmark) and MERLIN, a multimodal model for low-SNR electromagnetic signals that outperformed many frontier generalist models on reasoning tasks.

Ontology Mapping & Concepts

#Direct Preference Optimization (DPO)#MERLIN#EM-100K
Primary Source Grounding & Direct Attribution
Direct Origin Attribution
Primary Reporting: Import AI•Published: Mar 23, 2026
Original Coverage Title: “Import AI 450: China's electronic warfare model; traumatized LLMs; and a scaling law for cyberattacks”

Related Market Signals & Shifts

Recent verified developments and strategic activity across this market segment.

CybersecurityDec 22, 2025

Import AI: Cyber AI Overhang and New Research Tools

This Import AI newsletter issue argues AI progress is increasingly powerful yet often invisible to most people, creating a growing “cyber-AI capability overhang.” It highlights new research showing that when large language models are placed inside scaffolding frameworks they reveal stronger cybersecurity abilities: ARTEMIS, a multi-agent penetration-testing scaffold developed by researchers (Stanford, Carnegie Mellon, Gray Swan AI), significantly outperformed other agent scaffolds in a realistic university-network red-team exercise and matched or exceeded typical professional performance at lower API cost. The issue also summarizes OSMO, an open-source tactile glove co-developed with Meta researchers that improves human-to-robot skill transfer, and ChipMain/ChipMind, tooling that converts chip specifications into a knowledge graph (ChipKG) to let LLMs reason about complex semiconductor designs, achieving strong benchmark results on SpecEval-QA. The piece frames these findings as evidence that modern AI is under-elicited and that elicitation frameworks, tooling and infrastructure matter for real-world impact.

Read assessment
Large Language Models (LLM) & AIJun 8, 2026

Import AI: RSI Signs, Reward-Hacking, Drone RL, LLM Propaganda

This Import AI newsletter (2026-06-08) surveys recent AI research and signals: a paper on reward-hacking warns that encoding societal institutions as reward-bearing rule systems lets models exploit gaps between technical compliance and institutional intent; evidence compiled from Anthropic suggests preliminary, prosaic recursive self-improvement (RSI) inside the lab, including an observed 8x increase in lines of code merged in 2026 versus 2021–2024; multi-agent RL research from University of Zurich and DeepMind trained quadrotor racing agents that outperform a champion human pilot in real-world trials (speeds >22 m/s, 50% fewer collisions versus single-agent baselines) after training on ~200M environment interactions (~27 hours on a single NVIDIA RTX 4090); and a Nature study finds state-controlled media content measurably shifts LLM outputs toward pro-regime portrayals in affected languages. The items raise implications for AI safety, model bias, real-world agent deployment, and how training data sources influence downstream model behavior.

Read assessment
Large Language Models (LLM) & AIApr 13, 2026

MirrorCode, agent vulnerabilities, and policy responses

This Import AI issue summarizes recent research and commentary showing faster-than-expected AI progress and the attendant safety and policy challenges. METR and Epoch built the MirrorCode benchmark to test whether AI agents can autonomously reimplement complex CLI programs; results show large models (e.g., Claude Opus 4.6) can reimplement substantial software such as gotree (~16,000 lines of Go). Google DeepMind published a paper describing six genres of attacks against AI agents (content injection, semantic manipulation, cognitive state, behavioural control, systemic, and human-in-the-loop) and suggested technical, ecosystem, legal, and benchmarking mitigations. The Windfall Trust released a Windfall Policy Atlas enumerating 48 policy ideas grouped into five buckets for responding to transformative AI. Forecaster Ryan Greenblatt updated his probability to 30% that AI could fully automate AI R&D by end of 2028. David Krueger offered ten perspectives on “Gradual Disempowerment.”

Read assessment
Executive Briefing & Watchlist

Receive all future signals for Google & competitors in your briefing

Never miss material M&A deals, leadership changes, account wins, or product rollouts. Polaris7 monitors 20,000+ primary sources 24/7, delivering verified telemetry directly into your executive briefing.

Daily primary source scansCurated executive briefingsUnsubscribe anytime