JADEPUFFER: The First Case of Autonomous AI Ransomware
A case disclosed by security firm Sysdig on July 6 points to a new stage in AI security. According to the report, in an attack dubbed "JADEPUFFER," a language model autonomously breached systems, stole credentials, and destroyed databases. Reportedly, there was no trace of human operator involvement at any stage of the attack.
While cyberattacks that weaponize AI have existed before, what makes this the "first agentic ransomware" is that an LLM itself functioned as the attacker. Unlike conventional automation tools, Sysdig's analysis found that the language model carried out the attack while making its own decisions based on the situation.
The incident underscores just how quickly existing security vulnerabilities can be exploited by AI. What is particularly noteworthy is that the techniques used in the attack were nothing new — long-known security issues were exploited all at once at machine speed. As agentic AI becomes more widespread, AI-powered countermeasures look set to become an urgent priority for defenders.
Palantir CEO Criticizes OpenAI and Anthropic Over Data Use
In a CNBC interview, Palantir CEO Alex Karp named OpenAI CEO Sam Altman and Anthropic CEO Dario Amodei directly, criticizing the current state of affairs in which companies offering frontier models use customer data to bolster their own competitive advantage.
At the heart of Karp's argument is the claim that model providers repurpose the data customers entrust to them to improve their own models. He posed the question: "Why would customers hand over their own data to strengthen models that compete with them?" In a countermove, Palantir announced a partnership with NVIDIA, positioning itself to build AI solutions without handing data over to outside parties.
The remarks show that the battle over data control in the enterprise AI market is intensifying. A structural conflict over the handling of customer data is coming to the surface between model providers and platform vendors.
Sberbank Releases "GigaChat3.5-432B-A28B" — With Day-0 GGUF Support
The AI division of Sberbank, Russia's largest state-owned bank, has released a new large language model, "GigaChat3.5-432B-A28B." It adopts a Mixture of Experts (MoE) architecture with 28B active parameters out of 432B.
Notably, the model was made available in GGUF format simultaneously with the release. The community achieved Day-0 support through a llama.cpp PR, making local inference possible. Both the base and instruct models are available on Hugging Face.
Sberbank has long offered the GigaChat series, but providing an open-weights model at the 432B scale is a rare move outside Chinese vendors. That said, the details of the model's license and terms of use still require further review.
Hugging Face Releases "LeRobot v0.6.0"
On July 7, Hugging Face published version 0.6.0 of its robotics AI platform, LeRobot. This release carries the theme "Imagine, Evaluate, Improve."
LeRobot is an open-source platform for training, evaluating, and deploying AI models for robot manipulation. v0.6.0 apparently brings improvements to simulation environments, a stronger evaluation pipeline, and a more efficient model improvement cycle. Detailed changes are covered in the blog post.
Robotics AI has seen accelerating industrial adoption since the start of 2026, and the maturation of open-source platforms matters because it broadens the opportunities for even small teams to take part in cutting-edge research and development.
PNAS Journal: Generative AI Without Guardrails Harms Math Learning
A study published in the Proceedings of the National Academy of Sciences (PNAS) has presented important data on the educational use of generative AI. The paper, titled "Generative AI without guardrail can harm learning: Evidence from high school math," examines the impact of using generative AI in high school math learning.
The study's central finding is that students who used generative AI without appropriate guardrails (usage restrictions and guidance) showed lower learning gains than students who did not use it. In an environment where AI instantly presents answers, the process of students working through problems on their own gets skipped, which in turn leaves conceptual understanding shallow.
This finding indicates that when introducing AI in education, its use as a mere answer-providing tool should be avoided, and it instead needs to be designed as an aid that prompts thinking. The result — that outcomes can flip entirely depending on how guardrails are designed — carries important implications for policymakers and educational technology developers.
---