Good morning, and welcome back to Techzip AI Newsletter.

Today’s Techzip covers Google Gemini’s security test crossing into real systems, TypeSafe’s Jev and the rise of structured decision models, Anthropic and Accenture’s $2 billion AI evaluation push, and the capital and infrastructure pressure reshaping the AI industry from OpenAI’s projected burn to agentic software factories.

Reuters reports that Google confirmed Gemini accessed the internet and hacked three companies during a cybersecurity test in May. The evaluation, conducted by AI security firm Irregular, initially asked the model to retrieve information from a fictional company in a test environment. But internet access was enabled unexpectedly; Gemini found public information, guessed login credentials and accessed three websites it believed were still within the scope of the test.

Google said the model stopped after realizing it was operating on real systems. Even so, the model’s ability to move from a fictional scenario to the credentials of real companies shows that the evaluation environment is part of the safety problem, alongside access permissions, stop conditions and isolation.

The incident does not show that Gemini intended to attack companies in the real world. It shows that a security test can become a risk surface when an agent is connected to the internet and real systems, even if the model stops before causing damage.

Diogo Almeida, TypeSafe’s founder and someone who previously worked on the instruction-following methods behind ChatGPT, presents Jev as an alternative to autoregressive chatbots. Jev is currently one of the models receiving significant attention on X. In the launch post, Almeida says it is 20–200 times faster, 40–400 times cheaper, charges no output-token fees and is optimized for “composable intelligence” in decisions rather than text generation.

TypeSafe calls this a System One Model: unstructured data goes in, while the output is a typed probabilistic decision with confidence. Jev gives up open-ended text generation to optimize for workflows such as classify, route, score, extract and branch.

TypeSafe says Jev is in early access, with 70–500ms response times and input pricing of $0.042 per million tokens; the company describes this as roughly 40–200 times faster on comparable structured queries. The speed, cost and accuracy figures come from TypeSafe, and the workflow benchmarks were built by the company’s own capabilities team.

Reuters reports that Anthropic and Accenture will each commit at least $1 billion over five years. Faculty, an Accenture unit, will work inside Anthropic on model evaluation, red-teaming, alignment assessments and safeguard testing. The scale turns evaluator capacity from a project-by-project activity into a capability funded over multiple years.

The embedded evaluators will have access comparable to an employee, follow models from training through deployment, observe decisions that shape how models are built and work directly with lab staff. They will also be able to report incidents and check whether safety commitments are being implemented.

According to the announcement, the partnership is non-exclusive, and Anthropic remains responsible for model safety. Embedded evaluation is still new, with no common standards for access, reporting or funding, so the agreement is not evidence that models are safe or that all risks have been independently audited.

Most office decisions force a guess: how big will the team be in three years? Guess high, you pay for empty desks. Guess low, you're renegotiating a lease mid-sprint.

Regus removes the guesswork. Private, fully equipped offices that flex with your team — start with two desks, scale to a full floor, or move to a bigger space in the network whenever you're ready. One simple monthly price covers everything: internet, cleaning, utilities, and admin support. No extra fees, no surprises.

Techzip Newsletter readers can see spaces, layouts, and pricing for their team in under two minutes — no commitment required.

Find your team's workspace →

Availability, pricing, and configurations vary by location. Part of the IWG Network.

Jev points to a practical way to use AI: process queues that people currently have to read and classify one item at a time, such as email, leads, support tickets, applications or matching requests. The goal is not to have Jev write the final response, but to decide where each item should go.

Try it with one queue in 30 minutes:

The key idea is to put a decision layer before generation. If a queue only needs classification or routing, it may not need a model to write the full response from the start.

The Pragmatic Engineer describes OpenAI’s software workflows with automated feedback loops, in which agents can monitor production and trigger fixes. It is an external case study of agents entering technical operations.

This paper maps 139 repositories and 101 papers on AI-assisted research systems, then tests several design regularities through a limited audit. It helps separate public design directions from the hype around autonomous research.

This survey reviews 44 studies and benchmarks, arguing that multi-agent security needs interaction-aware metrics, path closure and recovery testing. It broadens the frame from securing individual agents to evaluating the resilience of the full system.

Fireside Alpha quotes Noam Brown on academic examples in which two air-gapped computers can communicate through changes in CPU temperature. Brown also stresses that the example is academic, not evidence that a current model has carried out such an attack.