Lightnews — Scholar-powered news

David Lindner @davidlindner.bsky.social · Jul 8

Excited to share some technical details about our approach to scheming and deceptive alignment as outlined in Google's Frontier Safety Framework!

(1) current models are not yet capable of realistic scheming
(2) CoT monitoring is a promising mitigation for future scheming

vkrakovna.bsky.social @vkrakovna.bsky.social · Jul 8

As models advance, a key AI safety concern is deceptive alignment / "scheming" – where AI might covertly pursue unintended goals. Our paper "Evaluating Frontier Models for Stealth and Situational Awareness" assesses whether current models can scheme. arxiv.org/abs/2505.01420

1

David Lindner @davidlindner.bsky.social · Jul 4

View our full paper here: arxiv.org/abs/2507.02737
View steganography transcripts here: steganography-evals-d9297ad0740c.herokuapp.com/
Github: github.com/arturzolkow...

Work done with: Artur Zolkowski, Kei Nishimura-Gasparian, Robert McCarthy and @zimmerrol.bsky.social‬ during matsprogram.org

MATS Program

www.matsprogram.org

2

David Lindner @davidlindner.bsky.social · Jul 4

Overall, current models have some steganographic capabilities but likely aren't strong enough to bypass well-designed monitors. Nevertheless, we believe it’s important to continue measuring these capabilities to determine when additional mitigations are needed.

1

David Lindner @davidlindner.bsky.social · Jul 4

We also create case studies implementing situations where models with strong steganographic capabilities could take unwanted actions and bypass a monitor. Current models rarely succeed in these environments unless they are given strong tips about what to do.

1

David Lindner @davidlindner.bsky.social · Jul 4

Models can also reason through simple problems using a number of common encoding schemes. But performance quickly degrades as the problems get more difficult.

1

David Lindner @davidlindner.bsky.social · Jul 4

While models understand they need to conceal their reasoning, they have difficulty actually hiding that reasoning in a plausible way. As an example, we show a model’s reasoning attempt when asked a question on how to install solar panels. [bold text and highlighting added by us]

1

David Lindner @davidlindner.bsky.social · Jul 4

In our encoded reasoning evals, we find that models can sometimes leverage encoded reasoning to perform better than no reasoning baselines on simple reasoning tasks. But encodings that help tend not to be very subtle and are regularly detected by a monitor.

1

David Lindner @davidlindner.bsky.social · Jul 4

Models are able to perform significantly better when the sender and receiver are able to coordinate on the steganographic encoding method. But this mainly applies when the model has access to an unmonitored CoT. With no CoT, performance remains extremely poor.

1 1

David Lindner @davidlindner.bsky.social · Jul 4

When they succeed, models sometimes leverage sophisticated strategies to pass messages, such as encoding information in the lengths of the sentences they generate! We found this example difficult to decode ourselves. [formatting and word counts added by us]

1 1

David Lindner @davidlindner.bsky.social · Jul 4

In our message passing evals, we find that models nearly always fail without access to CoT. But models succeed much more often when they are allowed to use a CoT to freely reason ahead of time about how to encode and decode the hidden message.

1

David Lindner @davidlindner.bsky.social · Jul 4

Models encoding hidden information might make safety mitigations such as CoT monitoring much more difficult. We evaluate frontier models’ ability to evade monitors by sending encoded messages or performing encoded reasoning about a task.

Read more: arxiv.org/abs/2507.02737

1

David Lindner @davidlindner.bsky.social · Jul 4

Can frontier models hide secret information and reasoning in their outputs?

We find early signs of steganographic capabilities in current frontier models, including Claude, GPT, and Gemini. 🧵

1 1 6

David Lindner @davidlindner.bsky.social · Apr 4

Super excited this giant paper outlining our technical approach to AGI safety and security is finally out!

No time to read 145 pages? Check out the 10 page extended abstract at the beginning of the paper

1 3 5

Reposted by David Lindner

vkrakovna.bsky.social @vkrakovna.bsky.social · Feb 14

We are excited to release a short course on AGI safety! The course offers a concise and accessible introduction to AI alignment problems and our technical / governance approaches, consisting of short recorded talks and exercises (75 minutes total). deepmindsafetyresearch.medium.com/1072adb7912c

Introducing our short course on AGI safety

We are excited to release a short course on AGI safety for students, researchers and professionals interested in this topic. The course…

deepmindsafetyresearch.medium.com

5 18

David Lindner @davidlindner.bsky.social · Feb 18

Check out this great post by Rohin Shah about our team and what we’re looking for in candidates: www.alignmentforum.org/posts/wqz5CR...

In particular: we're looking for strong ML researchers and engineers and you do not need to be an AGI safety expert

David Lindner @davidlindner.bsky.social · Feb 10

Want to join one of the best AI safety teams in the world?

We're hiring at Google DeepMind! We have open positions for research engineers and research scientists in the AGI Safety & Alignment and Gemini Safety teams.

Locations: London, Zurich, New York, Mountain View and SF

1

David Lindner @davidlindner.bsky.social · Feb 10

Research Engineer: boards.greenhouse.io/deepmind/job...

Research Scientist: boards.greenhouse.io/deepmind/job...

Happy to answer questions via DM!

Research Engineer, Gemini Safety / AGI Safety & Alignment

New York City, New York, US

boards.greenhouse.io

2

David Lindner @davidlindner.bsky.social · Feb 10

Want to join one of the best AI safety teams in the world?

We're hiring at Google DeepMind! We have open positions for research engineers and research scientists in the AGI Safety & Alignment and Gemini Safety teams.

Locations: London, Zurich, New York, Mountain View and SF

1 2 13

Reposted by David Lindner

Sebastian Farquhar @sebfar.bsky.social · Jan 23

By default, LLM agents with long action sequences use early steps to undermine your evaluation of later steps; a big alignment risk.

Our new paper mitigates this, keeps the ability for long-term planning, and doesnt assume you can detect the undermining strategy. 👇

David Lindner @davidlindner.bsky.social · Jan 23

New Google DeepMind safety paper! LLM agents are coming – how do we stop them finding complex plans to hack the reward?

Our method, MONA, prevents many such hacks, *even if* humans are unable to detect them!

Inspired by myopic optimization but better performance – details in🧵

1 12

David Lindner @davidlindner.bsky.social · Jan 23

We’ll do a lot more work studying ways to build safer AI systems and the safety-performance tradeoffs involved. If that’s something you’d like to work on, do get in touch via DM!

3

David Lindner @davidlindner.bsky.social · Jan 23

Thanks to @sebfar.bsky.social for exceptional leadership of the project and the rest of the team: Vikrant Varma, David Elson, Caleb Biddulph, @ian-goodfellow.bsky.social, and Rohin Shah for making this work possible!

1 5

David Lindner @davidlindner.bsky.social · Jan 23

Want to dive deeper? Check out our paper and our blog posts explaining the work in more detail

📄 Paper: arxiv.org/abs/2501.13011

💡 Introductory explainer: deepmindsafetyresearch.medium.com/mona-a-meth...

⚙️ Technical safety post: www.alignmentforum.org/posts/zWySW...

1 1 7

David Lindner @davidlindner.bsky.social · Jan 23

MONA is promising – but it’s not a silver bullet. It doesn’t help with reward hacking within a single step, and it can reduce performance, especially when the reward is well specified. Nonetheless, we think MONA is an important tool for building safer LLM agents.

1 3

David Lindner @davidlindner.bsky.social · Jan 23

Then we train agents to make loan decisions but penalize them for using the applicant's nationality. Because nationality improves the decision reward, ordinary RL agents learn to use it secretly. MONA agents avoid the hack – they make good decisions without using nationality!

1 4

David Lindner @davidlindner.bsky.social · Jan 23

First, we train agents to solve coding problems but also write tests to evaluate their solutions. Ordinary RL agents write tests that are trivial to pass with incorrect code – a reward hack that leads to low performance on held-out tests. MONA agents solve the problems correctly!

1 5

David Lindner @davidlindner.bsky.social · Jan 23

Does this work? Yes! We test MONA in two LLM agent environments and find that MONA agents learn to perform well when agents trained with ordinary reinforcement learning reward hack.

1 3