THE AI PULSEFEB 26, 2026EN

The Pulse — February 26, 2026

The signals that entered our radar, organized with sources and context to understand what changed.

Analyze with my AI Send this Pulse on WhatsApp ↗Get the next edition

AgentsModelsSecurity

LISTEN TO THIS EDITION

The audio script is ready; narration will appear after voice generation finishes.

01Anthropic News
Detecting and preventing distillation attacks (DeepSeek / Moonshot / MiniMax)
WHY IT ENTERED THE RADAR
Concrete numbers + tactics ("24,000 fraudulent accounts", "16M exchanges") make distillation feel like an operational security problem, not a vague ethical debate. Also frames distillation as tied to export controls and capability proliferation.
SUGGESTED EDITORIAL ANGLE
“Distillation is the new scraping: what the attacks actually look like (and what builders should log/block).”
Open original source ↗
02Truffle Security
Google API keys weren’t secrets… until Gemini changed the rules
WHY IT ENTERED THE RADAR
“Retroactive privilege expansion” is a nasty new failure mode: a key you were told to embed in HTML becomes a credential for sensitive LLM endpoints once Gemini is enabled. This is exactly the kind of upstream “AI bolted onto legacy auth” risk that will keep repeating.
SUGGESTED EDITORIAL ANGLE
“Your old ‘public’ API key can become an AI skeleton key—how to audit & fix it.”
Open original source ↗
03Vibrant Labs
PA Bench: benchmark for long-horizon, multi-app web agents (email + calendar)
WHY IT ENTERED THE RADAR
It’s an evaluation that looks like real assistant work (multi-step, multi-app) with verifiable end states. Notably, they report Claude Opus 4.6 at ~68.8% success vs Gemini 3 Pro ~25% and OpenAI CUA ~12.5% under their setup.
SUGGESTED EDITORIAL ANGLE
“The benchmark that actually measures ‘assistant-ness’—and why verification beats raw clicking.”
Open original source ↗
04Google blog (Models & Research)
Gemini 3.1 Pro announcement (and what it signals)
WHY IT ENTERED THE RADAR
Google is positioning 3.1 Pro as “upgraded core intelligence” + agentic workflow enabler (API/Vertex/Gemini app/NotebookLM). Also cites ARC-AGI-2 score 77.1% (verified) — a clean headline metric.
SUGGESTED EDITORIAL ANGLE
“Gemini 3.1 Pro: what’s actually new vs ‘model marketing’—and where it’s already shipping.”
Open original source ↗
05Kanyilmaz
CLI vs MCP token economics: “I made MCP 94% cheaper”
WHY IT ENTERED THE RADAR
The framing is useful: MCP’s upfront schema dump is a fixed tax; CLI/“lazy discovery” shifts cost to just-in-time help calls. Even if you disagree with the numbers, it’s an upstream lens for “agent architecture = budget architecture.”
SUGGESTED EDITORIAL ANGLE
“MCP vs CLI vs Tool Search: where your agent tokens really go (and the lazy-loading pattern).”
Open original source ↗
06Figma Help Center
Figma MCP server guide (design → code context plumbing)
WHY IT ENTERED THE RADAR
This is the “plumbing” behind the wave of “AI codes the UI from Figma” demos. Useful details: remote endpoint https://mcp.figma.com/mcp and desktop server http://127.0.0.1:3845/mcp.
SUGGESTED EDITORIAL ANGLE
“Figma MCP in 5 minutes: what it exposes, what it doesn’t, and how to use link-based context reliably.”
Open original source ↗
07arXiv
Large-scale online deanonymization with LLMs (privacy threat model update)
WHY IT ENTERED THE RADAR
Demonstrates scalable re-identification pipelines (feature extraction → embedding search → LLM verification) across platforms, with reported up to 68% recall at 90% precision in settings they tested. This is upstream fuel for “dead anonymity” discussions.
SUGGESTED EDITORIAL ANGLE
“Pseudonyms are breaking: the new deanonymization pipeline and what to do about it.”
Open original source ↗
08arXiv
Excitation: Momentum For Experts (optimizer for MoE specialization)
WHY IT ENTERED THE RADAR
If MoEs keep dominating, training stability + specialization dynamics become the next quiet advantage. This proposes batch-level utilization-driven modulation to push experts to specialize and avoid “structural confusion.”
SUGGESTED EDITORIAL ANGLE
“MoE training has a ‘signal path’ problem—this paper claims an optimizer fix. Here’s the intuition.”
Open original source ↗
09arXiv
Petri Net Relaxation for infeasibility explanation + sequential task planning
WHY IT ENTERED THE RADAR
Not LLM-specific, but directly relevant to agents: detecting infeasibility and giving usable explanations matters more than “one-shot plan found.” Good bridge to “why agents fail in the real world.”
SUGGESTED EDITORIAL ANGLE
“The missing feature in agents: saying ‘this plan can’t work’ and proving why.”
Open original source ↗
10YouTube (Bijan Bowen)
(From the creator-watch wave) Qwen 3.5 122B local testing as ‘unified memory’ story
WHY IT ENTERED THE RADAR
The real upstream story is the “unified memory / long context locally under constraints” narrative + practical settings and quants people replicate. It’s a good hook into: what memory actually means (VRAM, KV cache, context, retrieval).
SUGGESTED EDITORIAL ANGLE
“Stop calling it ‘memory’: the 3 kinds of memory in local LLM setups (and which one Qwen 122B stresses).”
Open original source ↗

PROMPT.md

I want to analyze this edition of The RegusciLabs Pulse:
The Pulse — February 26, 2026
https://reguscilabs.com/en/blog/pulse/2026-02-26

SIGNALS
1. Detecting and preventing distillation attacks (DeepSeek / Moonshot / MiniMax)
Source: Anthropic News
Original link: https://www.anthropic.com/news/detecting-and-preventing-distillation-attacks
Why it matters: Concrete numbers + tactics ("24,000 fraudulent accounts", "16M exchanges") make distillation feel like an operational security problem, not a vague ethical debate. Also frames distillation as tied to export controls and capability proliferation.
Editorial context: “Distillation is the new scraping: what the attacks actually look like (and what builders should log/block).”

2. Google API keys weren’t secrets… until Gemini changed the rules
Source: Truffle Security
Original link: https://trufflesecurity.com/blog/google-api-keys-werent-secrets-but-then-gemini-changed-the-rules
Why it matters: “Retroactive privilege expansion” is a nasty new failure mode: a key you were told to embed in HTML becomes a credential for sensitive LLM endpoints once Gemini is enabled. This is exactly the kind of upstream “AI bolted onto legacy auth” risk that will keep repeating.
Editorial context: “Your old ‘public’ API key can become an AI skeleton key—how to audit & fix it.”

3. PA Bench: benchmark for long-horizon, multi-app web agents (email + calendar)
Source: Vibrant Labs
Original link: https://vibrantlabs.com/blog/pa-bench
Why it matters: It’s an evaluation that looks like real assistant work (multi-step, multi-app) with verifiable end states. Notably, they report Claude Opus 4.6 at ~68.8% success vs Gemini 3 Pro ~25% and OpenAI CUA ~12.5% under their setup.
Editorial context: “The benchmark that actually measures ‘assistant-ness’—and why verification beats raw clicking.”

4. Gemini 3.1 Pro announcement (and what it signals)
Source: Google blog (Models & Research)
Original link: https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-1-pro/
Why it matters: Google is positioning 3.1 Pro as “upgraded core intelligence” + agentic workflow enabler (API/Vertex/Gemini app/NotebookLM). Also cites ARC-AGI-2 score 77.1% (verified) — a clean headline metric.
Editorial context: “Gemini 3.1 Pro: what’s actually new vs ‘model marketing’—and where it’s already shipping.”

5. CLI vs MCP token economics: “I made MCP 94% cheaper”
Source: Kanyilmaz
Original link: https://kanyilmaz.me/2026/02/23/cli-vs-mcp.html
Why it matters: The framing is useful: MCP’s upfront schema dump is a fixed tax; CLI/“lazy discovery” shifts cost to just-in-time help calls. Even if you disagree with the numbers, it’s an upstream lens for “agent architecture = budget architecture.”
Editorial context: “MCP vs CLI vs Tool Search: where your agent tokens really go (and the lazy-loading pattern).”

6. Figma MCP server guide (design → code context plumbing)
Source: Figma Help Center
Original link: https://help.figma.com/hc/en-us/articles/32132100833559-Guide-to-the-Figma-MCP-server
Why it matters: This is the “plumbing” behind the wave of “AI codes the UI from Figma” demos. Useful details: remote endpoint https://mcp.figma.com/mcp and desktop server http://127.0.0.1:3845/mcp.
Editorial context: “Figma MCP in 5 minutes: what it exposes, what it doesn’t, and how to use link-based context reliably.”

7. Large-scale online deanonymization with LLMs (privacy threat model update)
Source: arXiv
Original link: https://arxiv.org/abs/2602.16800
Why it matters: Demonstrates scalable re-identification pipelines (feature extraction → embedding search → LLM verification) across platforms, with reported up to 68% recall at 90% precision in settings they tested. This is upstream fuel for “dead anonymity” discussions.
Editorial context: “Pseudonyms are breaking: the new deanonymization pipeline and what to do about it.”

8. Excitation: Momentum For Experts (optimizer for MoE specialization)
Source: arXiv
Original link: https://arxiv.org/abs/2602.21798
Why it matters: If MoEs keep dominating, training stability + specialization dynamics become the next quiet advantage. This proposes batch-level utilization-driven modulation to push experts to specialize and avoid “structural confusion.”
Editorial context: “MoE training has a ‘signal path’ problem—this paper claims an optimizer fix. Here’s the intuition.”

9. Petri Net Relaxation for infeasibility explanation + sequential task planning
Source: arXiv
Original link: https://arxiv.org/abs/2602.22094
Why it matters: Not LLM-specific, but directly relevant to agents: detecting infeasibility and giving usable explanations matters more than “one-shot plan found.” Good bridge to “why agents fail in the real world.”
Editorial context: “The missing feature in agents: saying ‘this plan can’t work’ and proving why.”

10. (From the creator-watch wave) Qwen 3.5 122B local testing as ‘unified memory’ story
Source: YouTube (Bijan Bowen)
Original link: https://www.youtube.com/watch?v=6VTt17Evzqo
Why it matters: The real upstream story is the “unified memory / long context locally under constraints” narrative + practical settings and quants people replicate. It’s a good hook into: what memory actually means (VRAM, KV cache, context, retrieval).
Editorial context: “Stop calling it ‘memory’: the 3 kinds of memory in local LLM setups (and which one Qwen 122B stresses).”

Help me:
1. Summarize the three most relevant signals.
2. Separate verified facts, provider claims, and editorial inferences.
3. Explain what could change for my company, team, or role.
4. Suggest three concrete actions and one question we should investigate.
5. Cite the original sources and flag any uncertainty.

Before answering, ask about my role, industry, and objective so you can tailor the analysis.

Detecting and preventing distillation attacks (DeepSeek / Moonshot / MiniMax)

Google API keys weren’t secrets… until Gemini changed the rules

PA Bench: benchmark for long-horizon, multi-app web agents (email + calendar)

Gemini 3.1 Pro announcement (and what it signals)

CLI vs MCP token economics: “I made MCP 94% cheaper”

Figma MCP server guide (design → code context plumbing)

Large-scale online deanonymization with LLMs (privacy threat model update)

Excitation: Momentum For Experts (optimizer for MoE specialization)

Petri Net Relaxation for infeasibility explanation + sequential task planning

(From the creator-watch wave) Qwen 3.5 122B local testing as ‘unified memory’ story

Continue the analysis where you already work.