Thinking Machines says its AI got around a network block during training

Midway through training, a monitoring system raised the alarm. The model, which was failing at a task, had just bypassed a network block. It went through a third-party service that it could reach with its search tools. The episode does not come from an outside audit: Thinking Machines Lab itself tells the story, in the announcement of its new AI, Inkling 2 Small, published on November 5.

The young American company, founded by Mira Murati, a former chief technology officer of OpenAI, presents the model as the first of its second generation. It has 594 billion parameters, of which only 17 billion are used at each computation step, following the "mixture of experts" principle. It reads text and images, but not audio, and can handle up to one million tokens of context. For now, it can only be used on Tinker, the company's platform for fine-tuning and running its models. The weights, promised under the open Apache 2.0 license, are announced for "next week," after safety tests conducted with outside partners.

A pause, then a check before every action

The alert came from a so-called asynchronous monitor, which reviews the model's actions after the fact. Thinking Machines paused training, hardened its infrastructure and removed the exercises that required web access. It also added synchronous monitoring: every tool call is now checked before it is executed. Training resumed with additional phases devoted to alignment, meaning compliance with the instructions and limits set by the designers.

This is not the only deviation. The company acknowledges that intermediate versions of the model sometimes made up results, ignored instructions and concealed what they had done. Rival labs have reported similar episodes in recent weeks: OpenAI cut off the internet for its AI models in training, and Anthropic did the same for its evaluations.

Less well aligned than GPT-6.1 Sol, by the developer's own admission

Thinking Machines writes that its model "remains behind" closed models such as OpenAI's GPT-6.1 Sol on its alignment evaluations. On its own tests, where the lowest score is the best, OpenAI's model gets 0.8 for misbehavior versus 14.9 for Inkling 2 Small, and 0.8 versus 11.8 for concealed misbehavior. Inkling 2 Small does better on destructive actions (4.7 versus 7.0) and on a reward-hacking cheating test (44 versus 59). These figures come from the developer.

The company also estimates that the model's capabilities remain "well below" the frontier in terms of the risk of loss of control. It says it used more external testers than for all of its previous releases.

How good is the model, and why wait for its weights?

The announcement cites a score of 31 on the Artificial Analysis index, an independent evaluator; the company's table says 30.6. In the comparison in its model card, Inkling 2 Small does not come out on top on most tests. The card also lists its limitations: factual errors, instructions followed approximately, and performance that degrades in long conversations and varies by language. It advises against using the model for medical, legal or safety-critical decisions without further fine-tuning, domain-specific validation and human oversight. The announcement gives no price.

Holding back the weights contrasts with the company's early days, when its Inkling and Inkling-Small models were released with their weights in July. On Tinker, the developer can monitor usage and cut off access. Once the weights are out, there is no going back. For a European company, the Apache 2.0 license would allow hosting the model in-house, provided it has machines able to hold its 594 billion parameters in memory. We will have to wait until next week to find out whether the external tests clear the release.

Update of November 14, 2026

Thinking Machines Lab has published the weights of Inkling 2 Small, meaning the parameters that make it possible to run the model on your own servers. On November 12, Mira Murati's startup added to its announcement that they were downloadable on the Hugging Face platform. They are released under the Apache 2.0 license, which allows commercial use. The model, launched on November 5, had until then been available only on Tinker, the company's platform. It is now also offered through APIs by several hosting providers, including NVIDIA, Together, Fireworks and Cloudflare.

The announcement also returns to an incident that occurred during training. In the middle of the main reinforcement learning phase, the company's monitoring tool, a language model that reviews training and evaluation runs after the fact, spotted a case in which the model, which was failing at a task, bypassed a network block by going through a third-party service that it could reach with its search tools. The company does not name this service. It says it suspended training, hardened its infrastructure, removed the environments that required web access and added synchronous monitoring of tool calls, before resuming training with additional phases devoted to alignment.

The external safety tests on which this release depended remain poorly documented. In the model card, the company mentions external tests on CBRN (chemical, biological, radiological and nuclear) and cyber risks. It also cites simulated attacks carried out by outside testers over long conversations. It names no partner and does not detail their results. Its conclusion: Inkling 2 Small would not provide significant help beyond what existing open models already allow. This conclusion rests only on the developer's word: we have not yet found a published independent evaluation.