AGTP (@AGTPinsights) on X

What it is

An X thread by AGTP (@AGTPinsights) summarizing Meta's Muse Spark 1.3 Evaluation Report, the safety write-up for the first Meta model in the Muse family. Meta released Muse Spark 1.3 on September 2, 2026. Zifan (Sail) Wang announced the report. The thread is secondhand and quotes none of the report's prompt injection scores.

Prompt injection details

Meta used two new benchmarks from Exponential Security Labs, RepoInject v1 and BrowserInject v1. These appear to cover repository and browser agent settings, though the thread does not describe them. Exponential says Muse Spark 1.3 is about as robust as GPT-5.6 Sol.

Meta also ran Gray Swan's IPI Arena on the Q1 and Q2 splits (through April), the same splits other model cards used. That makes the result comparable with those cards. The thread says Meta reports "strong results" and gives no figures.

Maksym Andriushchenko of Exponential adds that prompt injection is not solved yet.

Other safety coverage

• Chain of thought controllability: Meta joins OpenAI, Anthropic and Google DeepMind in reporting it. Tomek Korbak says Muse Spark's controllability stays low, which suggests weak ability to evade monitors. Korbak wishes Meta had also reported compliance.
• Loss of Control: PostTrainBench is listed among the key catastrophic-risk evals.
• Human red teaming: Handshake tested before launch (section 3.1.4.2) with pre-release access. Jailbreaks were found through the product, and full trajectories went back to Meta to improve mitigations.

Related model facts in the thread

• Muse Spark 1.3 has a 1M-token context window, 75.4% on DeepSWE 1.1 and 98.5% on long-context MRCR.
• The open-weight Muse Spark Open has 816B total and 42B active parameters, with a 1M-token context window. Together AI hosted a live session on it, and Applied Compute put it live on AC2 for training and inference.

Caveats

• The "on par with GPT-5.6 Sol" claim comes from Exponential, and the thread gives no numbers. Check the report itself before relying on it.
• The Midas Project says Meta released Muse Wonder on Oct. 12 with a 13-page model card and a promise of a safety report soon. It says Meta updated the card 11 days later and there is still no safety report. This is the Midas Project's claim, relayed through the thread.