Gonzalo Gonzalez

Research Scientist @ Collinear AI | PhD @ Cornell | Building recursively self-improving agents

[Gonzalo Gonzalez @chalo2000]

@m_bourgon I think so. If Anthropic can invest in cyber and post-training guardrails to prevent distillation attacks from competitors then they should be able to make similar guardrails to prevent models from teaching themselves malicious behavior towards humans.

[Gonzalo Gonzalez @chalo2000]

There is quality control for vendor data that frontier labs buy to ensure their models don't go to shit; I'm sure you can add checks on model-generated data to avoid incentivizing killing humans (unless we want to benchmaxx that, in which case I'll share you samples)

[Evan Hubinger @EvanHub]

To be clear, as we say in our latest Risk Report (https://t.co/pG69KaI7a4), I think the risk from present models is low. What I am worried about is superintelligence arising from recursive self-improvement, as we have said is happening faster than we thought (https://t.co/aQoIG2eJHM).

[Sundar Pichai @sundarpichai]

We’re also introducing Gemini 3.8 Flash Cyber, our most capable cybersecurity model. It shows frontier-level performance in discovering vulnerabilities and patching them at scale, with Flash-level speed & pricing.

That includes achieving 86.2% on the important CyberGym industry benchmark, plus 47.2% on CWE-Bench for patching. We saw a 70%+ success rate in discovering vulnerabilities across 20 programming languages on our internal benchmark.

[Gonzalo Gonzalez @chalo2000]

was great working with @adit @chinmaya_mohan and everyone at @CollinearAI on this effort!

[Nazneen Rajani @nazneenrajani]

Today we’re releasing CWE-bench: 100 held-out audit-and-patch tasks testing whether coding agents can defend real code.

The leading agent passes 47%. 18 tasks remain unsolved by any agent.

This means frontier models on our benchmark have de-correlated errors.

Blog: https://t.co/7aMg0nypS6

[Gonzalo Gonzalez @chalo2000]

@nazneenrajani @OpenAI @huggingface The same capabilities leading to agents escaping sandboxes can be used to patch them to prevent this behavior fully. Sandbox escaping is a result of reward hacking from training on insecure envs and bad verifiers. A lot of work will need to be done to correct this misalignment.

[Nazneen Rajani @nazneenrajani]

Everyone has been talking about their agents escaping sandboxes since the July incident where two of @OpenAI models escaped a sandboxed cyber eval via a "read-only" registry proxy and ended up in @huggingface prod infra hunting for benchmark answer keys.

But rarely people talk about the environment or the simulated world around the agent.

I wrote a essay on it and propose a measurement for the environment's attack surface 🧵

Blog link: https://t.co/H4ZhaydpCs

[Ang Li @angli_ai]

Agent S3, the first computer-use agent that surpasses human-level performance in OSWorld v1 benchmark, has now been accepted to @TmlrOrg (Transactions on Machine Learning Research).

Congratulations to the @SimularAI team!

@chalo2000, vincent, @Richard_simular, jiachen, @xwang_lk

Stay tuned for what's next!

[Gonzalo Gonzalez @chalo2000]

@CollinearAI @anand_k27 @alckasoc

[Gonzalo Gonzalez @chalo2000]

@soumyadeepb_ @CollinearAI @anand_k27 @alckasoc Too busy eating the free chicken tenders

[Gonzalo Gonzalez @chalo2000]

@CollinearAI @alckasoc @icmlconf I had a great time at ICML learning about people's work on RSI related works. Happy to connect and chat with anyone working on RSI!

[Gonzalo Gonzalez @chalo2000]

@pkseeg @CollinearAI @alckasoc @icmlconf

[Collinear AI @CollinearAI]

Our researchers @chalo2000 and @alckasoc just returned from @icmlconf with a lot to share!

A takeaways 🧵 (1/8)

First we presented our work, YC-Bench, a long-horizon planning and consistent execution benchmark!

[Yunhai Feng @yunhaif]

Can robots learn contact-rich tool use, like operating a pair of scissors✂️ or turning a screwdriver🪛, from human data?

Introducing REGRIND: a minimalist retargeting-guided RL recipe for dexterous manipulation.

🌐 https://t.co/zzydolIgkc
📄 Paper + code on the site!
🧵👇

[Gonzalo Gonzalez @chalo2000]

@CollinearAI @aclmeeting @pkseeg @anand_k27 Super-human coding abilities but still really stupid when interacting with humans and remembering my preferences. Need a good user simulator to hillclimb on so I can trust my agent as much as my collaborators.

[Collinear AI @CollinearAI]

https://t.co/MiYMeUdOhD

[Maximus Pace @maxpace1_]

Teaching a robot shouldn't require humans to act like robots.

Human demonstrations contain valuable signal for robot manipulation, but they aren’t directly transferable to robots.

X-Diffusion learns from noisy human demonstrations while staying within the robot’s capabilities.

[Collinear AI @CollinearAI]

Most “hard” problems are useless for training a model.

The useful ones sit in a narrow learnable region where the model fails sometimes and succeeds sometimes.

Here are some early results where we took 5 trivial and 5 impossible cybersecurity environments and had a model rewrite them toward that region (green) over a few rounds. 🧵

1/7

[Gonzalo Gonzalez @chalo2000]

@CollinearAI

[Gonzalo Gonzalez @chalo2000]

sometimes I lose my Claude Code sessions, so I made a Claude Code session to find my Claude Code sessions

[Gonzalo Gonzalez @chalo2000]

@CollinearAI me rerunning fable to catch all my codebase vulnerabilities