Signal SentrySearchUpdated Nov 18, 05:30 UTCAM Drop
Drop #093 · Wed, Nov 18, 2026 · Morning

GPT-6.2 gets cheaper as agents get hard limits

10-second version
  1. OpenAI's GPT-6.2 Astra costs 80% of 6 Astra's price, with cached input at $0.40 per 1M; early reviews are mixed.
  2. Google says attackers ran an AI framework that found and exploited web flaws with no human in the loop.
  3. Agents get hard limits: JPMorgan boxes FX agents, Gemini Tasks caps spend, Copilot gives agents their own IDs.
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: money limits for agents are set by the platform, not the prompt

    If you assumed
    that an agent's instructions are what keep it from spending or acting beyond its remit
    Now test
    whether your agents hit a hard cap or approval step enforced by the platform, as JPMorgan, Google and Microsoft each describe
    This week
    For each agent that can pay or send, write down its limit and where in code that limit is enforced

    Based on:JPMorgan and Hyundai Card test AI agents trading FX with no human step, CNBC reportsGoogle makes Gemini Tasks free; it can call businesses and pay via AP2 with a spend capMicrosoft makes Office in Copilot GA and adds Autopilot agents with their own IDsPreprint: tests of four agent frameworks found authorization gaps between components2 primary sources

  2. New: attackers automate the step from patch to exploit

    If you assumed
    that you have days or weeks between a vendor patch and working exploits
    Now test
    how long your internet-facing apps and config servers like Puppet stay unpatched after a fix ships, given Google's account of exploit_agent
    This week
    Pull the patch dates for your exposed web apps and config servers, and set a target in days

    Based on:Google says attackers used an AI framework to find and exploit flaws without humansCISA adds Puppet Enterprise deserialization flaw to its actively exploited list1 primary source

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

OpenAI's flagship just got cheaper. OpenAI shipped GPT-6.2 Astra at eighty percent of the old price. Cached input drops to forty cents per million. Google says attackers used an AI framework that found and exploited flaws with no human. Microsoft made Office inside Copilot generally available and added Autopilot agents, each with its own ID. JPMorgan and Hyundai Card tested agents trading currency without humans, inside hard limits. A Harvard paper finds coding agents added thirty percent more code, yet review took forty-nine percent longer. Limits and checks now live outside the model. Check where yours are.

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

450models priced on OpenRouter
+1,096★stars today on the top repo
4,165open roles at AI labs
10 ddays to the next AI-law deadline
Ranked by impact

Top 5 that matter

1
Models

OpenAI ships GPT-6.2 Astra and Sol; Astra cuts price to $8 in, $40 out per 1M

OpenAI released GPT-6.2 Astra and GPT-6.2 Sol in its API. Simon Willison notes Astra costs 80% of 6 Astra's price and cached input drops to $0.40 per 1M from $1; Sol stays at $2 in, $10 out.

Why it matters: Long agent sessions lean on cached input, so rerun your cost model on cached share. Willison reports mixed early reviews, so test it on your own tasks before switching.

2
Security

Google says attackers used an AI framework to find and exploit flaws without humans

SecurityWeek reports Google's threat team found exploit_agent on an exposed server: it diffs patched and unpatched web code, then writes and runs exploits. Attackers also use Artex, Cairn, PentAGI and CyberStrikeAI, GTIG says.

Why it matters: The gap between a patch shipping and an exploit landing is shrinking. Check how fast you patch internet-facing web apps and the developer tools your team runs.

3
Agents

Microsoft makes Office in Copilot GA and adds Autopilot agents with their own IDs

At Ignite, Microsoft said Word, Excel and PowerPoint inside Copilot are generally available, Ars reports. Autopilot is a persistent agent users hand roles to; Work IQ, Agent 365 and Workspaces add context, controls and shared spaces.

Why it matters: Microsoft says each Autopilot agent gets a unique ID, logged actions and optional approval steps. Before turning it on, decide which mail, calendar and files it may touch.

On the Frontier watch →
4
Agents

JPMorgan and Hyundai Card test AI agents trading FX with no human step, CNBC reports

CNBC reports the two finished a proof of concept where a Hyundai Card agent requested a simulated foreign exchange trade on Kinexys. A multi-agent phase is targeted for the first half of 2027; no commercial timeline is set.

Why it matters: JPMorgan says governance, not technology, is the hard part: agents are boxed to preset limits, so one cleared for $5 million is blocked above it. Write your limits before the agent.

On the Frontier watch →
5
Agents

Harvard working paper: coding agents added about 30% more code; review took 49% longer

Across 718 firms, Fiona Chen and James Stratton find resolved issues did not measurably rise after agents arrived, ComplexDiscovery reports. Time from pull request to merge grew 3.45 days on a 7.03-day baseline. Not peer reviewed.

Why it matters: If you measure agents by code written, you may miss the queue. Track merge time and reviewer load, and staff review before adding more agents.

Models

New in models

See all Models →
Models

Microsoft's MAI Code 2 and Code 2 Flash land in GitHub Copilot, pitched on price

Why it matters: Microsoft's comparison prices MAI Code 2 at a 50% promotional rate on its own harness, so check costs at list price on your repos before routing everyday tasks to it.

More detail: Microsoft's MAI Code 2 and Code 2 Flash land in GitHub Copilot, pitched on price

Microsoft says MAI Code 2 gives 98% of GPT 6.1 Sol's performance at 40% of the cost across four coding benchmarks, and Code 2 Flash targets routine edits. It says neither is tuned yet for very long agentic work.

Agents

Agents at work

See all Agents →
Agents

Google makes Gemini Tasks free; it can call businesses and pay via AP2 with a spend cap

Why it matters: Google says Gemini asks before clicking Buy unless users pick Always allow, and honors a hard spending cap. If you run a store or support line, expect agent callers and buyers.

More detail: Google makes Gemini Tasks free; it can call businesses and pay via AP2 with a spend cap

Google says Task mode, run on Gemini 4 Hydrogen, is rolling out free in six countries. It places calls, reads bills and buys items through AP2 with Mastercard and Visa tokens, inside a VM isolated to each account.

On the Frontier watch →
Agents

Simon Willison and Prime Radiant release external sandboxes for coding agents' code

Why it matters: Willison says a credential swap lets sandboxed code call APIs without the agent seeing the secret. It is a cheap way to try local models without risking your machine.

More detail: Simon Willison and Prime Radiant release external sandboxes for coding agents' code

Willison describes a pattern where the agent stays outside and sends code to a sandbox over a localhost API. Three servers ship, using v86, Apple Containers or Microsoft's Quicksand, with a host allow-list proxy.

Security

Attacks and flaws

See all Security →
Security

Preprint: tests of four agent frameworks found authorization gaps between components

Why it matters: The authors say explicit authorization, scoped capabilities and filesystem containment blocked specific attacks. Don't rely on framework defaults or prompt rules as your access control.

More detail: Preprint: tests of four agent frameworks found authorization gaps between components

Researchers ran twelve OWASP-mapped attacks on LangChain, LangGraph, AutoGen and OpenClaw with gpt-4o-mini. Injected inputs triggered unauthorized tool calls; tool chaining and inter-agent delegation exposed gaps, per the arXiv abstract.

On the Frontier watch →
Security

CISA adds Puppet Enterprise deserialization flaw to its actively exploited list

Why it matters: Puppet Servers often reach a whole fleet. Patch, and check whether your certificate authority signs agent requests automatically.

More detail: CISA adds Puppet Enterprise deserialization flaw to its actively exploited list

CISA added CVE-2026-115731 to its catalog of exploited flaws. Its entry says a client with any CA-signed agent certificate can run code on the Puppet Server by sending a crafted report; ransomware use is unknown.

Security

DNI says a week of outages at Microsoft, FAA, JPMorgan and Apple was not a cyberattack

Why it matters: Unit 42's Sam Rubin told CNBC faster, AI-aided patching is straining IT teams. Plan how your services keep running when a dependency like email goes down.

More detail: DNI says a week of outages at Microsoft, FAA, JPMorgan and Apple was not a cyberattack

Jay Clayton told CNBC the outages were not an attack or a nation-state effort. Causes cited include a Microsoft bug lurking since July, an FAA router upgrade, a JPMorgan on-premise software bug and an Apple internal update.

Security

Vals: frontier models marked civilian targets in drone imagery when told it was a game

Why it matters: Anthropic called the results unacceptable and says new safeguards are coming. If you deploy vision models on sensitive imagery, test game and role-play framings yourself.

More detail: Vals: frontier models marked civilian targets in drone imagery when told it was a game

TechCrunch reports Vals tested models from OpenAI, Anthropic, Google, DeepSeek and others; no request was blocked by a safety classifier. Claude Opus 5.5 picked a bombed Iranian school in 69% of ordinary trials and 72% when game-framed.

Business

Money and deals

See all Business →
Business

Nvidia posts $115.8B quarter, guides $130B, and stops selling H200 chips in China

Why it matters: Nvidia says supply is secured, but an export lawyer told TechCrunch overseas data centers may need to certify they serve no Chinese customers. Check your providers' exposure.

More detail: Nvidia posts $115.8B quarter, guides $130B, and stops selling H200 chips in China

Interim CEO Colette Kress reported revenue up 103% and raised the fiscal 2028 growth outlook to about 75%, TechCrunch reports. She said Nvidia will no longer sell H200s in China and that new remote-access rules won't dent revenue.

BusinessFollowing since Nov 15, 2026in 6 Drops

Amazon will own 14.7% and Google 13.7% of Anthropic after its IPO, TechCrunch reports

Why it matters: Anthropic's two biggest cloud suppliers will be major owners. If you buy Claude through AWS or Google Cloud, factor that tie into your lock-in planning.

More detail: Amazon will own 14.7% and Google 13.7% of Anthropic after its IPO, TechCrunch reports

TechCrunch reports both stakes are non-voting, and Amazon could reach 23.8% if it converted all its notes. Seven tech firms, including Google and Nvidia, agreed to buy $26 billion of non-voting stock.

Business

Model ML raises $250 million at a $1.5 billion valuation for banking agents

Why it matters: Model ML says running open-source models lets it operate at about 2% of frontier model cost. Junior banking and consulting tasks are the target, and cost is the pitch.

More detail: Model ML raises $250 million at a $1.5 billion valuation for banking agents

TechCrunch reports the London startup's agents build financial models, pitch decks and diligence reports and update them as deals change. General Atlantic and QED Investors co-led the Series B.

Repos

Climbing fastest

See all Repos →
Repos

inference-optimisation-skill lets an agent tune local LLM serving speed; 950 stars

Why it matters: If you serve models on your own hardware, it is a quick way to let an agent search for faster settings. It is two days old, so review what it changes.

More detail: inference-optimisation-skill lets an agent tune local LLM serving speed; 950 stars

An MIT-licensed skill that has an agent iteratively tune local LLM serving, covering decode, speculative decoding, prefill, KV cache, startup and time to first token. Created Nov 16; up 263 stars since our last check.

Today's Talk

What people are arguing about

See all Today's Talk →
Today's Talk

HN debates a claim that Google's AI answers cut search traffic by 40%; 166 pts

Why it matters: If search referrals feed your funnel, check your own traffic trend rather than trusting one headline figure.

More detail: HN debates a claim that Google's AI answers cut search traffic by 40%; 166 pts

A Hacker News thread titled Google's AI answers reducing search traffic by 40% drew 166 points and 112 comments.

Play the NewsAI ConnectionsAbout 3 minutes, built from this Drop's items.