Google uses DeepMind AI to analyze Signal code for security flaws

Anthropic is developing a technology called FlashProve, designed to allow verification of whether a response was generated by the artificial intelligence (AI) model that a company claims to be using. This system employs cryptography to generate a type of mathematical proof that attests to the identity of the model responsible for a given output.

The main purpose of this proposal is to offer users the ability to confirm whether they are interacting with the AI advertised by the provider, and not with an inferior or different version. According to Anthropic, this mechanism can also make it difficult for companies to secretly swap their models for more economical options or present performance metrics that do not match the actual technology used.

The project was disclosed in a company publication on Tuesday (3). Furthermore, the company made a technical specification and parts of the code available in open format, inviting other researchers to analyze and improve the initiative.

One of the central problems that FlashProve seeks to solve is the situation where a provider presents one model to clients but operates with another during service use. For example, an organization might advertise access to advanced AI but replace it with a lower-cost alternative to reduce expenses. The mechanism would also help prevent cases where a company uses a different model than the one it declared using when proving a result.

The applications of this verification go beyond the direct services of AI companies. Researchers cite, for example, a developer who claims to have used an advanced cybersecurity model to inspect a software library for vulnerabilities. Another potential application would be a company stating that its business plan received approval from an AI suitable for that task.

With FlashProve, it would be possible to perform a cryptographic check to confirm whether the advertised model was actually used in these processes. Despite the public disclosure, FlashProve remains a research project. Anthropic released a detailed technical specification and functional code of various components under an open-source license, allowing the community to examine and contribute to improving the system.

However, the company has not announced plans to integrate this technology into its own services or other AI platforms. Researchers clarified that the current publication represents an intermediate stage of the work, not the completion of the entire research plan. The intention is to enable other teams to evaluate the proposal, use its components, and help refine the tool.

Google succeeded in running an artificial intelligence (AI) model on a satellite, receiving a response generated from outside Earth. This test marks an experimental advancement of Project Suncatcher, a corporate initiative focused on exploring the feasibility of transporting AI data centers into Earth orbit.

In a statement released on the company's blog, Google detailed that it began using Tensor Processing Units (TPUs) installed on the space equipment. During the tests, the Gemma 4 model was run directly on the satellite and received an instruction to compose a haiku about the experience of being in space.

The AI generated poems addressing Earth viewed from space, data flow, and the brilliance of the stars. Furthermore, other experiments were conducted, such as creating limericks with space themes and responding to telemetry-related commands.

Additionally, Google is working to improve communication between devices operating in orbit. After the initial testing phases, the company plans to achieve a new goal in 2027: deploying two interconnected satellites.

The purpose of this goal is to validate whether the systems are capable of performing tasks consistently and exchanging data in the space environment. Such a demonstration would be crucial in determining the practicality of establishing an orbital computing infrastructure.

Currently, the test conducted with Gemma 4 only confirms that an AI model can process commands aboard a satellite and produce responses. The realization of space data centers still requires the completion of several technical stages and the execution of more experiments.