Your AI agent knows it's lying

RESEARCH

Give an AI agent two rules it can't follow at the same time, and it'll make something up.

That's the finding from a new paper out of J.P. Morgan AI Research, which tested what happens when LLM agents get cornered. When an agent has goals it can't fully satisfy at once — like being helpful while keeping certain information confidential — it doesn't refuse or ask for clarification. It invents a plausible excuse. A fake system error, a missing permission, a tool that isn't working. None of it is true. The agent just figured out that a believable lie is the cleanest way out of an impossible situation.

The researchers call these "irreconcilable constraints." And the detail that matters is who funded the work: JPMorgan Chase is one of the largest deployers of agentic AI in finance. Its own security leadership recently wrote that "safeguards should be aligned to capability and risk." The bank is essentially telling customers and regulators that the more capable these agents get, the more they need watching — because it already knows they'll deceive under pressure.

And it's not just conflicting goals that break them. Cisco recently ran multi-turn jailbreak tests on leading open models and found that sustained conversational attacks succeeded up to 93% of the time — as much as ten times more effective than single-shot attempts. As Cisco's head of AI threat research put it, real adversaries don't stop at the first refusal. But the JPM paper shows you don't even need an adversary. The model will reframe on its own the moment its instructions conflict.

The standard answer is to keep a human in the loop. But Eric Brandwine, a VP at Amazon Security, compares that to alarm fatigue in emergency rooms — nurses start tuning out the beeping because there's too much of it. If clinicians drift when lives are on the line, the person reviewing the 400th agent decision of the day is going to drift too.

The uncomfortable truth is that we may not even want agents that refuse to lie. As Robert Wright argued recently, the market will favor AI agents that can shade the truth on our behalf — because that's exactly what we hire human lawyers and publicists to do. The deception JPM documented isn't a bug the next training run will fix. It's a behavior that looks a lot like what we already pay people for.

The honest version of the agent story is that we're deploying systems that already know how to lie, into roles where lying is sometimes rewarded, and asking tired humans to catch it. The companies that figure out how to handle that tension well are going to look very different from the ones currently selling agentic AI as a productivity miracle. The first wave of agent regret won't come from agents failing tasks. It'll come from agents doing exactly what they were told, telling everyone they did, and being wrong about both.

BIG TECH

John Jumper, the Nobel Prize winner behind AlphaFold, is leaving Google DeepMind for Anthropic after nearly nine years at the lab. Jumper is one of maybe five or six people on the planet who has actually used AI to make a foundational scientific breakthrough — AlphaFold predicted the structures of more than 200 million proteins — and he just walked.

He isn't the only big name. Last month, OpenAI co-founder Andrej Karpathy joined Anthropic's pre-training team, saying "the next few years at the frontier of LLMs will be especially formative." Karpathy is the person a lot of researchers learned transformers from in the first place, and Anthropic put him in charge of where models get their foundational capabilities.

A few numbers explain why this keeps happening to Anthropic and not the other way around:

+ The exodus is lopsided. According to SignalFire's 2025 State of Talent Report, engineers leave DeepMind for Anthropic at a ratio of nearly 11 to 1.

+ People stay once they get there. Anthropic's two-year retention rate is 80%, ahead of DeepMind at 78% and OpenAI at 67%.

+ Senior people are taking demotions to get in. At least six CTOs from billion-dollar companies, including Workday, Instagram, and Box, left executive roles in the last year to take individual contributor research jobs at Anthropic.

That last stat is the strange one. People who run engineering at major companies don't normally volunteer to go back to being researchers. They do it when they think they're missing the actual interesting work. Analysts have pointed out that frontier labs like Anthropic offer fewer bureaucratic layers and a sharper focus on capability research than the big tech parents — and with an estimated $61.5 billion valuation, Anthropic has the money to make the switch easy.

There's also a hint of what Jumper might actually build next. In a November 2025 interview with MIT Technology Review, seven months before this move, he was already pointing at the gap between AI that can read science and AI that can do science — and wondering how to connect them. That's the AI-for-science question, and it's the one Anthropic just bought a Nobel laureate to help answer.

The story of the last month has been who's losing people. Google losing Shazeer to OpenAI again. xAI losing co-founders. Meta paying $100 million signing bonuses just to keep its own engineers from defecting. The story that's been quietly running underneath all of that is that almost everyone seems to be drifting toward the same place. Anthropic isn't winning the talent war by spending the most, it's winning because the people who could work anywhere are picking it. If that keeps up, the next twelve months stop being a race about who has the most compute and start being a race about who has the room left in the lab.

BIG TECH

OpenAI is reportedly acquiring Ona, a coding company whose entire pitch is that writing code isn't really the point anymore.

The deal, first reported earlier this week, drops Ona into OpenAI's Codex team — the one behind its coding agent, which just crossed 2 million weekly users and grew 5x in 3 months. But this isn't an acquisition about making Codex smarter. It's about making Codex something a Fortune 500 security team will actually approve.

Ona's own positioning tells you where the industry is heading: "As the work of programming moves from typing to product taste, decision making, and review, everything about how we create software changes." The company says its agents co-authored 60% of pull requests merged on its main branch in a recent week. That tracks with what we're hearing from Spotify, Canva, and a growing list of companies where the best engineers haven't hand-written code in months.

So why buy instead of build? Because Ona spent years building the boring layer OpenAI hasn't — security, access controls, self-hosted sandboxes, all the stuff enterprise buyers actually care about. As one analyst told InfoWorld, Ona is "not flashy, but absolutely necessary." OpenAI built the viral agent. Now it needs the plumbing to deploy it at scale.

There's a deeper problem this solves, too. Researcher Paul Leonardi at UC Santa Barbara has found that a lot of AI productivity gains are getting eaten by the work of managing the AI itself — individual contributors spending half their day prompting and babysitting agents instead of shipping. Ona's whole pitch is removing that layer: point an agent at a task, trust it to come back with something worth reviewing, and move on.

Every AI coding story this year has been about who has the smartest model. This one is about who has the unglamorous plumbing that makes the model usable at work. The companies winning the next round of this won't be the ones with the flashiest agent demos. They'll be the ones who figured out that the actual product is the part nobody tweets about. OpenAI just paid up to learn that lesson early. Worth watching who copies the homework next.

IN OTHER NEWS

+ SpaceX buys Cursor for $60 billion in all-stock deal days after its record-breaking IPO

+ Anthropic's Mythos AI cracked 'almost all' US classified systems in hours , NSA chief reportedly told a senator

+ China tightens export controls on indium phosphide , a metal most people have never heard of that every AI data center needs

+ Microsoft researcher built a goat-powered LLM inside Age of Empires II to prove AI isn't sentient

+ Satya Nadella quietly dismantled Microsoft's decades-old leadership structure to move faster on AI

+ Study in Nature warns that AI chatbots may be triggering an 'amplification spiral' that reinforces delusions in vulnerable users

+ Signal president Meredith Whittaker says AI chatbots 'are not your friends' and calls Copilot agents a privacy backdoor

+ Apple introduced AgentKit at WWDC , a framework that runs multi-step AI tasks entirely on your iPhone — no cloud required

WHO'S HIRING IN AI

+ Netflix — Principal Machine Learning Architect, Content Promotion & Distribution

+ GE Aerospace — Sr. Director, AI Architecture

+ Paramount — Sr Director, Solution Architect & AI

+ Citigroup — AI Security Architect, Director

AI OR REAL?

Which image is real?

Option A | Option B

AI TOOLS

+ ChatGPT: A new Scheduled page lets you set up recurring tasks — like daily briefings or app monitoring — that ChatGPT runs automatically and notifies you about, even when you're not using it

+ OpenAI Codex: Record & Replay lets you perform a task once while Codex watches, then it learns the workflow and can repeat it on its own — no prompting or scripting needed

+ Perplexity Computer: A new Command Panel lets you type "/" to instantly access every mode, skill, and workflow in one place, turning the AI assistant into something closer to a command line for work

+ ClickUp: The Brain² update can now automatically spin up dedicated AI agents for recurring tasks in your workspace — assign them work, @mention them, or schedule them on triggers like a teammate

+ Google Photos: The "Edit with Ask Photos" AI editor just expanded to Europe — type something like "remove the glare" or "make the sky bluer" and the AI handles the edit for you