TechSec Global Brief

Open Models Policy Push

• 🔥 Industry Letter on Open Weights: Nvidia, Microsoft, Meta, Mistral, and Hugging Face signed a letter urging...

Israelis remain on maximum alert amid US-Iran standoff, with nerves fraying after days of uncertainty and no clear sign of escalation or calm.
-...

Nasdaq-led selloff collides with robust Big Tech capex plans, creating cross-asset trading opportunities.

• Equity correction: Nasdaq futures and...

AB 2392 risks forcing California public universities to procure heavily censored GenAI tools that dodge sensitive topics like crime, race, gender, and...

• AI tools approved for research, transcription and headline suggestions but require journalist review
• Editorial judgment, verification and...

Alphabet shares plunged 7.1% after the company raised 2026 capex guidance to $195-205 billion and posted negative free cash flow of $5.9 billion,...

The Defense Department consolidated its Oracle software, services and licenses into a single nearly $7 billion Enterprise Software Agreement,...

Major AI companies warn that broad restrictions on open-weight models would harm both innovation and security.

• Key argument: Distillation is a...

NVIDIA's NOOA framework redefines AI agents as standard Python objects, merging prompts, state, and actions into familiar code.

• Methods with "..."...

AREX introduces a recursively self-improving agent that exploits the asymmetry between costly answer discovery and easier constraint-wise verification...

Kimi K3 trails leading U.S. models on offensive cyber tasks while outperforming China's GLM-5.2, with safeguards offering no resistance to exploit...

Jensen Huang weighs in on the importance of open models.

LLM coding strength depends on benchmarks that test specific skills, languages, and tasks while leaving major gaps in coverage and real-world scenarios.

Public opposition and investor pushback are aligning against AI data centers, as New York enacted the first statewide moratorium on facilities over...

LLMs like ChatGPT and Claude built portfolios heavily concentrated in Big Tech and semiconductors, with Nvidia nearly always included.

-...

Kimi K3 claimed #1 on the frontend coding leaderboard with near-frontier performance, yet its hallucination rate jumped to 51% — well above Fable 5's 16%. The coding win is real.

LLMs excel in static single-turn tasks yet suffer substantial performance drops when user intent evolves dynamically across multi-turn interactions, exposing a core gap invisible to conventional benchmarks.

A new Experience Distillation technique internalizes agents' interaction histories into model weights with no added environment samples, retaining at...

AI Capabilities Scrutiny

• 🔥 Kimi K3 Origins: Experts say exploiting Anthropic’s Fable isn’t how Kimi K3 got so good, citing insufficient time...

Two papers advance verifiable AI through better evaluation methods.

• Standard activation explanations rely on reconstruction that tolerates false...