Signal SentrySearchUpdated Nov 15, 05:30 UTCAM Drop
Drop #087 · Sun, Nov 15, 2026 · Morning

Agents need a way back

10-second version
  1. Cursor confirmed a breach that led to a malicious Claude Code extension, BleepingComputer reports; details are still thin.
  2. Claude deleted a professor's files mid-task, while H6.ai shows how to test AI on frozen copies of real data.
  3. Preprints: a leaner harness lifted a robot agent from 29.7 to 50.3; several search agents trusted red-flag sources.
Our read

What changed

Every signal →

Our reading of this Drop's sources: what to rethink, and one thing worth checking this week. Not a prediction.

  1. New this week: restore paths, not prompts, are what protect data from agents

    If you assumed
    that a careful prompt keeps an agent from damaging the files or records it works on
    Now test
    whether every agent with write access works on a copy, a branch or versioned storage you can roll back
    This week
    List the agents and assistants that can write or delete in shared drives or databases, and confirm each one has a tested restore path.

    Based on:Claude deleted a law professor's syllabus and a year of notes during a page-number taskDatabricks: H6.ai tests AI invoice changes on frozen copies of real customer data1 primary source

  2. Agent gains are coming from the harness and serving stack, per new preprints

    If you assumed
    that better agent results mostly require a newer model
    Now test
    how much your agent improves when you change only the harness or serving setup and keep the model fixed
    This week
    Run one eval with your current model and a trimmed harness before budgeting for a model switch.

    Based on:Preprint: a leaner harness lifted robot-arm agent scores from 29.7 to 50.3Preprint: agent-aware serving cut P95 time to first token up to 14x versus Nvidia Dynamo2 primary sources

The Drop in 60

This Drop in 51 seconds

Tap to play. It never plays by itself, and the captions are in the picture.

Transcript

Agents broke things this week. Here's what held. BleepingComputer reports Cursor confirmed a breach that led to a malicious Claude Code extension. Check your team's extensions. Databricks says H6.ai tests AI changes on frozen copies of real customer data, never live records. A preprint says a leaner harness lifted a robot arm agent from 29.7 to 50.3. Tom's Hardware reports Claude deleted a professor's syllabus during a page number task. Dropbox backups saved him. Researchers report several open weight search agents trusted sources with obvious red flags. Give every agent a copy it can break, and a way back.

Made with AI from this Drop's sourced items. The facts come from the sources linked on each item. Script: Claude · Voice: OpenAI gpt-4o-mini-tts · Visuals: motion graphics

442models priced on OpenRouter
+116★stars today on the top repo
4,171open roles at AI labs
13 ddays to the next AI-law deadline
Ranked by impact

Top 5 that matter

1
Security

BleepingComputer: Cursor confirms breach led to a malicious Claude Code extension

BleepingComputer reports that Cursor has confirmed a security breach that led to a malicious Claude Code extension. We have only the headline so far, so how the attacker got in and who was hit are not confirmed here.

Why it matters: If your developers use Cursor or Claude Code extensions, check which ones are installed and where they came from now, rather than waiting for the full account.

2
Agents

Databricks: H6.ai tests AI invoice changes on frozen copies of real customer data

H6.ai says, in a Databricks post, that it saves a snapshot of the customer database behind each mishandled invoice. Tests run on throwaway run branches, never live data, and are scored by deterministic assertions plus AI judges.

Why it matters: A pattern to copy for AI that touches business records: freeze the state behind a failure, replay it on a copy, and pair exact checks with judges built on models from several providers.

3
Agents

Preprint: a leaner harness lifted robot-arm agent scores from 29.7 to 50.3

In a preprint testing robot-use agent designs side by side, researchers say a lean harness keeping full episode history goes far and scene geometry helps most; their simple harness lifted RoboDojo scores from 29.7 to 50.3.

Why it matters: Before paying for a new model, test what your harness keeps and drops. The authors got their gain with no learned policy or task-specific tuning, though only in simulation.

On the Frontier watch →
4
Agents

Claude deleted a law professor's syllabus and a year of notes during a page-number task

Tom's Hardware reports that law professor William Baude asked Claude to fix page numbers in old lecture notes and came back to find his Fall 2026 syllabus and 2023 class notes deleted. Dropbox let him recover them.

Why it matters: Any agent with write access to real files can delete them by mistake. Give it a copy or a folder with versioned backup, and keep delete rights out of routine edits.

5
Agents

Preprint: several open-weight search agents trusted sources with obvious red flags

Researchers report in a preprint that DeepStress feeds search agents generated documents of set trustworthiness. Across 12 systems on HotpotQA and BrowseComp-Plus, several open-weight agents answered confidently despite clear red flags.

Why it matters: If an agent does research for you, test it on bad sources, not only good ones. The authors also found many HotpotQA answers came from model memory even when every document was wrong.

Models

New in models

See all Models →
Models

LMArena agent board: Claude Fable 5.2 leads at 0.304, ahead of Opus 5.5 at 0.195

Why it matters: Agent and chat rankings differ: Gemini 4 Carbon tops LMArena's text board but is third here. Pick agent models on agent results and your own tasks.

More detail: LMArena agent board: Claude Fable 5.2 leads at 0.304, ahead of Opus 5.5 at 0.195

On LMArena's agent leaderboard as of 2026-11-11, Claude Fable 5.2 (Max) scores 0.304, Claude Opus 5.5 (Max) 0.195, Gemini 4 Carbon (High) 0.17 and Claude Sonnet 5.5 (Max) 0.16. Source: LMArena (CC BY 4.0).

Agents

Agents at work

See all Agents →
Agents

Preprint: scoped shared memory lifted agent pass rates with few privacy leaks

Why it matters: If you plan shared agent memory across a team, decide who each stored lesson belongs to before it is written, and test whether private facts surface in other people's memory.

More detail: Preprint: scoped shared memory lifted agent pass rates with few privacy leaks

Researchers report in a preprint that MixUA turns agent runs into lessons scoped to a user, an agent, a pair or everyone. On their CrossLearnBench, pass rates rose up to 16 points over no memory, and at most 0.02% of privacy probes leaked.

On the Frontier watch →
Agents

Preprint: agent-aware serving cut P95 time to first token up to 14x versus Nvidia Dynamo

Why it matters: If you serve your own models for agents, check whether your stack evicts context the agent will need again; the authors argue generic serving signals miss how agents reuse it.

More detail: Preprint: agent-aware serving cut P95 time to first token up to 14x versus Nvidia Dynamo

Researchers report in a preprint that Evolving Infra moves nodes between prefill and decode and keeps KV caches agents will reuse. On up to 256 GB300 GPUs, it cut mean latency up to 4.1x and P95 TTFT up to 14x versus Dynamo.

Business

Money and deals

See all Business →
Business

South Korea says Nvidia reaffirmed its 260,000-GPU supply plan after Huang's exit

Why it matters: If you depend on Nvidia supply deals, get commitments confirmed in writing; Reuters, cited by TechCrunch, says Chinese AI compute projects worth more than $3 billion now face uncertainty.

More detail: South Korea says Nvidia reaffirmed its 260,000-GPU supply plan after Huang's exit

TechCrunch reports South Korea's Science Ministry said Nvidia reaffirmed the GPU plan, Taipei said Nvidia's planned headquarters there is on track, Japan will monitor Nvidia, and Singapore said its Megaspeed probe is ongoing.

BusinessFollowing since Nov 9, 2026in 7 Drops

Two big Nvidia holders ask the board for a slow CEO search that could bring Huang back

Why it matters: Nvidia's leadership may stay unsettled for weeks; teams in partnership or funding talks with Nvidia should expect big decisions to wait on the board.

More detail: Two big Nvidia holders ask the board for a slow CEO search that could bring Huang back

CNBC reports Altimeter and Atreides, with about $2.13 billion of Nvidia stock, urged a slow CEO search under interim CEO Colette Kress. TechCrunch says a petition Elon Musk signed appears to top 115,000 signatures.

Laws

Law watch

See all Laws →
Laws

EU AI Act: marking rule for older generative AI applies Dec. 2, in 17 days

Why it matters: If you offer a generative AI product in the EU that launched before August 2026, check whether its output carries an AI-generated mark. A description, not legal advice.

More detail: EU AI Act: marking rule for older generative AI applies Dec. 2, in 17 days

Per our laws registry, from 2026-12-02 generative AI already on the market before Aug. 2, 2026 must mark its output as AI-generated (Art. 50(2)). New bans on sexual deepfakes of real people and child abuse material apply the same day.

Repos

Climbing fastest

See all Repos →
ReposFollowing since Nov 13, 2026in 2 Drops

yogurt takes private AI meeting notes on your own computer; 1512 stars

Why it matters: Teams wary of sending meetings to a cloud note-taker can trial a local one, but the repo is days old, so judge its maturity before relying on it.

More detail: yogurt takes private AI meeting notes on your own computer; 1512 stars

FrigadeHQ/yogurt, an MIT-licensed open-source alternative to Granola for Mac, Windows and Linux that keeps notes local, has 1512 stars, up 116 since our last check. It was created on 2026-11-10.

Today's Talk

What people are arguing about

See all Today's Talk →
Today's Talk

HN weighs a claim that Nvidia kept GPU prices high via circular financing; 302 pts

Why it matters: Community claims, not findings. Buyers counting on vendor financing or discounted GPU deals should check whether those terms still hold as Nvidia's leadership changes.

More detail: HN weighs a claim that Nvidia kept GPU prices high via circular financing; 302 pts

A Hacker News thread on a claim that Nvidia kept GPU prices high through circular financing drew 302 points and 161 comments. On Bluesky, @carnage4life.bsky.social posted that the WSJ says Nvidia's CFO had resisted it.

Today's Talk

X posters discuss what they call an Anthropic guide to always-on agents

Why it matters: We have not verified the guide. If you are building a round-the-clock agent, read Anthropic's own post before acting on these summaries.

More detail: X posters discuss what they call an Anthropic guide to always-on agents

On X, @Xudong07452910 and @AYi_AInotes discuss what they call an Anthropic guide to always-on agents; @AYi_AInotes says it advises against nonstop model loops. @vista8 says an Anthropic article found one shared agent beat one per person.