Signal SentrySearchUpdated Nov 3, 17:30 UTCPM Drop
Drop #064 · Tue, Nov 3, 2026 · Evening

Security vendors move in on AI agents

10-second version
  1. CrowdStrike buys agent monitoring and attack-testing firms and adds a kill switch for misbehaving AI agents.
  2. OpenAI raises bug bounties as the UK's NCSC warns of AI attacks on critical infrastructure by 2028.
  3. Apple will cap server-side Siri AI use; a CoworkOS preprint shows an agent running multi-round model training.
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. Defenders are budgeting for AI-speed flaw finding

    If you assumed
    that a yearly pentest and a modest bug bounty keep pace with attackers using AI
    Now test
    whether your own agents and MCP servers have been deliberately attacked by your tests, not only reviewed
    This week
    Run one prompt-injection and tool-misuse test against a production agent and the MCP servers it calls

    Based on:CrowdStrike buys agent security firms Fiddler and Dreadnode, adds an AI kill switchOpenAI raises bug bounties, with critical ChatGPT flaws now paying up to $30,000UK cyber chief expects AI attacks on critical infrastructure by 20282 primary sources

  2. Cloud AI use is getting capped and costly

    If you assumed
    that AI features bundled into a device or plan come with open-ended use
    Now test
    which of your team's AI tools cap server-side requests, and what breaks in a workflow when a cap hits
    This week
    Price your heaviest agent workflow at current token rates and set a monthly cap per person

    Based on:Apple says server-side Siri AI features will have usage limits, with paid extra accessYC's Garry Tan says he spends $100,000 to $200,000 a year on agent tokens

The Drop in 60

This Drop in 52 seconds

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

Transcript

Security is coming for your AI agents. CrowdStrike is buying Fiddler and Dreadnode, for about 495 million dollars, the Journal reports. It also added an agent kill switch. OpenAI says it more than quadrupled bug bounty midpoints. Critical ChatGPT flaws now pay up to 30,000 dollars. Britain's cyber chief expects AI attacks on critical infrastructure by 2028. Nvidia open-sourced Auto Ontology for agents. Without Vault, it stores database passwords unencrypted. And a CoworkOS preprint reports an agent post-trained a model across five rounds. If agents do your work, attack-test them, and keep a way to stop them.

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

424models priced on OpenRouter
4,193open roles at AI labs
25 ddays to the next AI-law deadline
Ranked by impact

Top 5 that matter

1
Business

CrowdStrike buys agent security firms Fiddler and Dreadnode, adds an AI kill switch

CrowdStrike agreed to buy Fiddler, which monitors agents, and Dreadnode, which attack-tests models and MCP servers; the WSJ puts the pair at about $495 million, SecurityWeek reports. A new kill switch halts a misbehaving agent.

Why it matters: Agent monitoring, attack testing and a stop button now come from a mainstream security vendor. Check whether you can halt any production agent and cut off its network access today.

SecurityWeek · Nov 3
2
Security

OpenAI raises bug bounties, with critical ChatGPT flaws now paying up to $30,000

OpenAI says it is more than quadrupling average reward midpoints for medium to critical bugs, with exceptional finds worth up to $172,000 with bonuses. A new internal sprint will pair its teams with Cathedral's researchers.

Why it matters: If you build on ChatGPT, Codex or the API, watch OpenAI's security notices: it is paying outside researchers far more to find flaws in those products.

OpenAI · Nov 2OpenAI · Nov 2
3
Security

UK cyber chief expects AI attacks on critical infrastructure by 2028

NCSC head Richard Horne said the UK will likely face AI-enabled attacks on critical infrastructure by 2028, AP reports via SecurityWeek. The NCSC report says AI will mostly scale existing tactics rather than find brand-new flaws.

Why it matters: The NCSC's advice is basic: patch devices, limit admin access, back up data. Old systems with known flaws in your stack become cheaper for attackers to find and hit.

SecurityWeek · Nov 3NCSC UK on X · Nov 3
4
Agents

Nvidia open-sources Auto Ontology, a data-meaning layer agents query over MCP

Nvidia's Apache-2.0 library builds an ontology from enterprise metadata and checks agent-written SQL against certified definitions, per its README. Without Vault, UI-added database credentials are stored unencrypted in Postgres.

Why it matters: Nvidia's pitch is that agents need trusted metric definitions, not just database access. If you try it, configure Vault, set token expiry and keep its internal API port private.

GitHub · NVIDIA/auto-ontologyHacker News · Nov 3
5
Agents

Preprint: CoworkOS harness ran five rounds of automated post-training on a Qwen model

Researchers report CoworkOS splits strategy from execution, with resumable workspaces and persistent memory. With GLM-5.3 driving it, the average score of Qwen3.6-35B-A3B on eight benchmarks rose from 33.19 to 42.95 over five rounds.

Why it matters: The authors show a harness that keeps state across interruptions carrying multi-round training work. Treat it as a pattern to test, with a person approving any model it produces.

On the Frontier watch →
Hugging Face Papers · Nov 2
Agents

Agents at work

See all Agents →
Agents

YC's Garry Tan says he spends $100,000 to $200,000 a year on agent tokens

Why it matters: A heavy user puts real numbers on agent cost and data access. Before copying this setup, decide which records an agent may read and who checks that scope.

More detail: YC's Garry Tan says he spends $100,000 to $200,000 a year on agent tokens

Tan told Walter Isaacson's podcast his agent has read his email, Dropbox and Apple Notes, and that YC runs an internal agent, Stray Light, holding context on every founder it has backed, TechCrunch reports. He says costs must fall.

TechCrunch · Nov 3
Agents

Preprint: CUA-Master-27B computer-use agent scores 80.2 on OSWorld Verified

Why it matters: Scores swing widely across suites in the same paper. When judging a computer-use agent, ask for results on several benchmarks, not one headline figure.

More detail: Preprint: CUA-Master-27B computer-use agent scores 80.2 on OSWorld Verified

Researchers report a two-loop framework that mines real workflows from instructional videos to build training tasks, with snapshot rollback for retries. Their 27B model scores 36.9 on OSWorld 2.0 and 25.0 on Agents' Last Exam.

Hugging Face Papers · Nov 2
Agents

NATO's Crossed Swords exercise adds an AI agent track and uses OpenAI's Daybreak

Why it matters: Military cyber teams are practicing work beside agents under time pressure. Builders of security agents should be ready to show buyers how a person stays in control.

More detail: NATO's Crossed Swords exercise adds an AI agent track and uses OpenAI's Daybreak

Over 250 participants from more than 40 countries are training in Tallinn, and for the first time the exercise uses OpenAI's Daybreak alongside locally hosted AI, SecurityWeek reports. The exercise director stresses human oversight.

SecurityWeek · Nov 3
Business

Money and deals

See all Business →
Business

Apple says server-side Siri AI features will have usage limits, with paid extra access

Why it matters: Even Apple is metering cloud AI rather than giving it away. If your product leans on platform AI features, plan for usage caps and pricing tiers.

More detail: Apple says server-side Siri AI features will have usage limits, with paid extra access

On Apple's earnings call, CEO John Ternus said features using server models will be capped and more capacity sold later, The Verge reports. Apple spent $2.2 billion on capex last quarter versus Amazon's $77.3 billion.

The Verge · Nov 3
Business

Apple will require 18+ ratings for apps with dating or AI romantic companion features

Why it matters: AI companion apps with romantic content lose teen users on iOS. Audit your app's features against the new descriptors before November 16.

More detail: Apple will require 18+ ratings for apps with dating or AI romantic companion features

Apple told developers that apps with dating or romantic relationship features need an 18+ rating, with new content descriptors required from November 16, TechCrunch reports. Apps must also flag sharing a user's real-world location.

TechCrunch · Nov 3
Business

OpenAI pilots a Grid Protectors Program and starts calling data centers compute campuses

Why it matters: Local power politics now shape where AI capacity gets built. If you depend on new US capacity, watch siting news for grid deals and demand-response terms.

More detail: OpenAI pilots a Grid Protectors Program and starts calling data centers compute campuses

OpenAI says its planned Effingham County, Georgia site could free up to 1,000 megawatts during extreme weather or emergencies, The Verge reports. In the announcement, compute campus outnumbered data center, which appeared twice.

OpenAI · Nov 2The Verge · Nov 2The Verge · Nov 3
Business

Trajectory raises $45 million at a $300 million valuation for continual model training

Why it matters: Post-training from your own feedback is becoming a product, not a research team. If a general model keeps making the same domain mistakes, price this route against doing it in-house.

More detail: Trajectory raises $45 million at a $300 million valuation for continual model training

Sequoia led the Series A for the 18-person startup, whose platform retrains open-source models from feedback on specialized tasks, TechCrunch reports. Harvey and Rogo already use it, and the platform is now publicly available.

TechCrunch · Nov 2
Laws

Law watch

See all Laws →
LawsFollowing since Nov 2, 2026in 2 Drops

EO 14434 definition of Super Intelligence is due Nov 28; agency use of SI is uneven

Why it matters: If you sell to US agencies, check how new solicitations word AI requirements, and watch for the proposed definition later this month.

More detail: EO 14434 definition of Super Intelligence is due Nov 28; agency use of SI is uneven

The Federal Register text sets a deadline 60 days after signing for proposed legislative language defining Super Intelligence. MIT Technology Review finds an ICE solicitation uses SI, while an FTC release and HHS have not adopted it.

Federal Register · Oct 2MIT Technology Review · Nov 3
Repos

Climbing fastest

See all Repos →
Repos

skein runs fail-closed swarm orchestration with epoch-fenced work; 259 stars

Why it matters: If parallel agents in your stack can deliver stale work after a task has moved on, this repo's design is a short read on one way to refuse it.

More detail: skein runs fail-closed swarm orchestration with epoch-fenced work; 259 stars

Balanov/skein describes fail-closed agent swarm orchestration that fences work by epoch and rejects late completions. MIT license, created Nov 2, up 2 stars since our last check.

GitHub · Balanov/skein
ReposFollowing since Oct 30, 2026in 8 Drops

Mika runs an on-device Android agent that operates apps and the web; 5966 stars

Why it matters: Phone agents that drive other apps keep drawing interest. Test what one can reach on a spare device before giving it real accounts.

More detail: Mika runs an on-device Android agent that operates apps and the web; 5966 stars

wangzan101/Mika, Apache-2.0 and created Oct 29, gained 76 stars since our last check, about 1179 a day by its listing.

GitHub · wangzan101/Mika
Today's Talk

What people are arguing about

See all Today's Talk →
Today's Talk

HN post says OpenAI may be quietly re-enabling training on Codex chats

Why it matters: If your team uses Codex on private code, check its data-training settings yourself rather than assuming an earlier choice still holds.

More detail: HN post says OpenAI may be quietly re-enabling training on Codex chats

A Hacker News post titled as a claim that OpenAI is silently re-enabling model training on Codex chats drew 25 points and 8 comments. We have not verified the claim.

Hacker News · Nov 3
Today's TalkFollowing since Nov 1, 2026in 4 Drops

Bluesky posters say Claude's game-clone limits also block legitimate reverse-engineering

Why it matters: If you use Claude Code for decompiling or compatibility work, test those tasks now and keep a fallback tool in case they are refused.

More detail: Bluesky posters say Claude's game-clone limits also block legitimate reverse-engineering

Posted on Bluesky by @carnage4life.bsky.social: developers report Claude Code refusing legitimate reverse-engineering work under the new rule. A post on the change by @culturecrave.co drew 137 likes.

Bluesky · @carnage4life · Nov 3Bluesky · @culturecrave.co · Nov 3
Play the NewsShip It or Block ItAbout 2 minutes, built from this Drop's items.