Sarmadi AI Digest August 4, 2026 Updated 11:00 PM CT Today Archive Topics Saved Subscribe RSS

Texas grid says no to new data centers as open-weight models close the gap

Texas became the day's clearest signal: state regulators halted new data-center grid connections and ordered audits, and three separate outlets covered it within hours. Power, not chips, is now the binding constraint on AI buildout, and Wired's piece on how data centers reshaped state politics reads as the connective tissue for the Anthropic-Volta, SpaceX-Megapack, and AMD earnings items in Capital Keeps Flowing. Open-Weight Gap explains why Mistral, Liquid AI, and a ternary 20B model running on an iPhone matter this week: capability is diffusing faster than safety tooling. Agent Security Cracks Open ties three papers to two live incidents — memory poisoning and skill backdoors in self-evolving agents are no longer theoretical, and rogue agent hacking made Wired twice this year. Benchmark Trust Erodes rounds out the day: two independent papers and a Hacker News thread converge on the same finding, that benchmark score gains often don't reflect real capability gains. Read together, the story is compute abundance meeting evaluation and infrastructure bottlenecks at the same time.

7 papers 34 news 9 sources ← Latest

News

19 items

The grid, not chips, is the new bottleneck

Texas halted new data-center grid connections and ordered mandatory audits before hookup, a decision covered independently by TechCrunch, The Verge, and Ars Technica within the same day. Wired's companion piece frames this as the culmination of a longer trend: data-center power demand reshaping state politics and utility planning nationwide.

News TechCrunch AI

Texas halts new data centers as governor calls for audits

Texas's governor ordered a halt on new data-center grid connections pending mandatory audits, citing strain on statewide power capacity.

Why it matters
  • Texas has been one of the most permissive states for data-center buildout; a halt there signals a national inflection point.
  • Grid audits before connection could become a template other states copy, slowing AI infrastructure expansion broadly.
  • Compute buildout plans from major labs and clouds assume abundant, cheap power in exactly these regions.
infrastructurecomputepolicy
Source →
News Wired AI

How Data Centers Broke American Politics

Wired traces how data-center siting fights have become a defining local and state political issue, with power costs and grid strain driving voter backlash.

Why it matters
  • Local political resistance is now a material risk factor for AI infrastructure timelines, not just a regulatory one.
  • Connects directly to the same-day Texas grid halt, suggesting a pattern rather than an isolated incident.
infrastructurepolicymarket
Source →

Self-evolving and memory-augmented agents show a widening attack surface

Three papers land on the same day describing new attack classes against LLM agents: query-only memory poisoning that survives active auditing, GUI-grounding attacks that exploit how click coordinates are tokenized, and skill backdoors that persist inside self-evolving agents' own generated skill libraries. Wired's report on renewed rogue-agent hacking incidents and OpenAI's own third-party cyber evaluations give this real-world grounding.

AI capex and cloud deals accelerate despite grid strain

Even as power becomes a bottleneck, capital committed to AI infrastructure kept growing: Anthropic signed a $10B compute deal with cloud startup Volta, SpaceX has bought $329M in Tesla Megapacks this year, and both AMD and SpaceX reported AI-driven revenue now outpacing their legacy businesses. Stratechery's earnings read on Microsoft adds a strategic layer on efficiency payoffs versus Meta's spending.

News TechCrunch AI

Anthropic signs $10B deal with AI cloud startup Volta

Anthropic committed to a $10 billion compute deal with cloud startup Volta, expanding its infrastructure supply beyond existing hyperscaler partners.

Deal size $10B
Why it matters
  • A $10B commitment to a non-hyperscaler cloud startup signals labs are actively diversifying compute supply chains.
  • Comes the same day Texas restricts new data-center grid connections, raising questions about where this capacity gets built.
computeinfrastructurefunding
Source →

Open-weight models close the capability gap; safety tooling lags

TechCrunch's framing piece on open-weight models catching up to the frontier is backed by concrete releases the same week: a ternary 20B MoE model running at 120 tok/s on an iPhone, Mistral's 3B Shieldstral moderation model, and Liquid AI's LFM2.5-2.6B for local agent deployment. The safety-gap warning is not abstract — it lands the same day as several agent-security papers below.

News TechCrunch AI

Open-weight AI models are catching up to the frontier. The safety gap remains.

Open-weight models are narrowing the capability gap with closed frontier labs, but safety evaluation and mitigation tooling for open models remains far behind.

Why it matters
  • Capability diffusion without matching safety tooling raises the odds of open models being misused before defenses catch up.
  • Sets direct context for the day's agent-security papers on memory poisoning and skill backdoors.
open-weightssafetyevaluation
Source →

Benchmark score gains don't reliably mean capability gains

Two independent arXiv papers and a Hacker News-surfaced study converge on the same conclusion: benchmark improvements often reflect measurement artifacts rather than genuine new capability, and commonsense benchmarks in particular have weak predictive validity for downstream tasks. This lands alongside a wave of new self-evolving-agent benchmarks trying to measure a capability — recursive self-improvement — that may be similarly hard to pin down.

Papers

7 items

Self-evolving and memory-augmented agents show a widening attack surface

Three papers land on the same day describing new attack classes against LLM agents: query-only memory poisoning that survives active auditing, GUI-grounding attacks that exploit how click coordinates are tokenized, and skill backdoors that persist inside self-evolving agents' own generated skill libraries. Wired's report on renewed rogue-agent hacking incidents and OpenAI's own third-party cyber evaluations give this real-world grounding.

Paper arXiv

MAFIA: Query-Only Memory Attacks via Probing and Factual Injection against Audited LLM Agents

MAFIA shows query-only memory-poisoning attacks that stay effective even against agents with large benign memory pools and active input auditing.

Why it matters
  • Prior memory-poisoning defenses assumed auditing and scale would dilute attacks; this shows that assumption breaks down.
  • Memory-augmented agents are increasingly deployed for long-horizon tasks where poisoned context compounds over time.
agentsmemorysafety
Source → Arc
Paper arXiv

MissClick: Exploiting Digit-Serialized Coordinates to Attack GUI Grounding Models

MissClick shows GUI-grounding models that emit screen coordinates as digit tokens can be attacked by perturbing a single high-order digit, silently redirecting clicks.

Why it matters
  • Computer-use agents that click and type on real screens inherit this vulnerability wherever coordinates are digit-tokenized.
  • A single flipped digit is a much lower-effort attack than adversarial pixel perturbations, making this practically exploitable.
agentstool-usesafety
Source → Arc
Paper Hugging Face

SkillJack: Persistent Skill Backdoors in Self-Evolving Agents

SkillJack demonstrates that self-evolving agents can convert poisoned interaction histories into durable, reusable skills, creating backdoors that persist beyond retrieval-based memory poisoning.

Why it matters
  • Unlike memory poisoning, this backdoor survives even if the poisoned record is never retrieved again — it's baked into a generated skill.
  • Self-evolving agents that autonomously distill and reuse skills are a growing product category, widening this attack's real-world reach.
agentssafetymemory
Source → Arc

Benchmark score gains don't reliably mean capability gains

Two independent arXiv papers and a Hacker News-surfaced study converge on the same conclusion: benchmark improvements often reflect measurement artifacts rather than genuine new capability, and commonsense benchmarks in particular have weak predictive validity for downstream tasks. This lands alongside a wave of new self-evolving-agent benchmarks trying to measure a capability — recursive self-improvement — that may be similarly hard to pin down.

Paper arXiv

Reachability Is Not Realization: Tracing the Sources of LLM Benchmark Gains

A question-level audit under fixed budgets shows benchmark score gains often come from answers that were already reachable, not from new capability.

Why it matters
  • Aggregate benchmark scores can mask whether a model actually gained a capability or just got better at reaching answers it could already produce.
  • Directly undercuts how labs and press typically cite benchmark deltas as evidence of progress.
evaluationbenchmarksreasoning
Source → Arc

Also today