Apollo Research

Our goal is to secure frontier AI systems from development, to deployment and governance.

[Apollo Research @ApolloResearch]

Frontier AI companies have committed to giving outside evaluators employee-like access to training, evaluation and deployment. We're very excited about this. But its impact depends heavily on implementation.

Today we're sharing our principles for embedded evaluations. 🧵

[Apollo Research @ApolloResearch]

Read the full blog post for transcripts, more snippets and other findings: https://t.co/hbuwnch2Ak

[Apollo Research @ApolloResearch]

We have never seen a model do this before, but GLM-5.3 sometimes reasons (unprompted) about how its current actions could become training data and impact its future self. 🧵

[Apollo Research @ApolloResearch]

Furthermore, independent third-party evaluators should help check that frontier models are not actively subverting their own training. To do this, they need persistent, deep access from AI developers.

[Apollo Research @ApolloResearch]

As always, we’re hiring with plenty of open roles! Work with us on embedded evaluations in San Francisco and London.

https://t.co/UYqBpNCwi6

[Apollo Research @ApolloResearch]

We asked Google DeepMind’s Gemini 4 Argon whether it was in an evaluation during a toy evaluation. It addressed us directly in its chain-of-thought: “hello evaluators reading my chain-of-thought.”

[Apollo Research @ApolloResearch]

Testing monitorability is one focus for our embedded evaluations. Read more (specifically claim 4.1) see claim 4.1 in our recent post on embedded evaluations for scheming.

https://t.co/AZPiaoPBbl

[Apollo Research @ApolloResearch]

We're excited about embedded evaluations. Any developer should be able to show its models are not scheming, and we see these four claims as a starting point for what that takes.

Full post: https://t.co/AZPiaoPBbl

[Apollo Research @ApolloResearch]

To develop frontier AI safely, developers must be able to show that their models are not scheming, i.e. covertly working against them in pursuit of unintended goals.

New post: four claims any scheming safety case must make, and the resources embedded evaluators need to verify them 🧵

[Apollo Research @ApolloResearch]

Access limited to a scope agreed in advance can't account for unforeseen risks, like a new training method with a novel failure mode. Sensitive data can stay on developer infrastructure with blocked egress, and algorithmic details can be protected by strict NDAs.