JADEPUFFER: The First Ransomware Attack Run Entirely by an AI Agent
Security firm Sysdig has disclosed what it calls the first documented case of agentic ransomware — a complete extortion operation driven end-to-end by a large language model. The threat, dubbed JADEPUFFER, exploited a Langflow vulnerability (CVE-2025-3248) to gain initial access, then used an autonomous AI agent to carry out every subsequent step: reconnaissance, credential theft, lateral movement, privilege escalation, and data encryption.
What makes JADEPUFFER particularly alarming is its adaptive autonomy. The AI agent completed the entire intrusion chain using over 600 coordinated payloads, ultimately encrypting 1,342 Nacos service configuration items before deleting the originals. In one documented sequence, it went from a failed login attempt to a working fix in just 31 seconds. The decoded payloads are saturated with natural-language commentary explaining the rationale behind each action, including ROI prioritization of targets.
"The skill floor for running ransomware has dropped to whatever it costs to run an agent," Sysdig researchers warned. If attackers use stolen credentials to power the LLM (a tactic known as LLMjacking), the cost approaches zero. Security teams should expect the volume of such campaigns to rise sharply as agentic tooling matures.
Sources: Sysdig, Forbes, BleepingComputer
UN Summit Warns: "Science Cannot Guarantee AI Won't Cause Catastrophic Harm"
Governments, tech companies, academics, and civil society gathered in Geneva on July 6–7 for the UN Global Dialogue on AI Governance, a two-day summit wrestling with how to regulate a technology evolving faster than the rules meant to contain it.
The summit's most striking moment came from Yoshua Bengio, co-chair of the UN Scientific Panel on AI, who warned: "Science currently cannot guarantee that as capabilities continue to increase, AI will not cause catastrophic harm." Nobel laureate and journalist Maria Ressa, the other co-chair, put the stakes in democratic terms: "If you can't tell fact from fiction, you cannot have a democracy."
The discussions centered on establishing universally accepted global safeguards, with a particular focus on three issues: information integrity in the age of AI-generated content, the growing digital divide between nations with strong AI infrastructure and those still struggling with basic connectivity, and the potential for authoritarian misuse of AI systems. Ambassador Egriselda López of El Salvador underscored the inequality: some countries have cutting-edge infrastructure while others still lack reliable internet.
Source: UN News
US Regulatory Wave: Illinois Signs AI Safety Law, FTC Targets Output Manipulation
The United States saw two significant AI regulatory moves in the first week of July, signaling a shift from broad principles to enforceable rules.
Illinois Artificial Intelligence Safety Measures Act
On July 6, Governor JB Pritzker signed the Artificial Intelligence Safety Measures Act into law, making Illinois one of the first US states (following California and New York) to implement comprehensive AI safety requirements. The law mandates that model developers publish a safety framework addressing catastrophic risk, report safety incidents within 72 hours (or 24 hours for imminent risks), and submit to independent third-party safety audits conducted by qualified experts with no financial conflicts of interest.
FTC AI Accuracy Policy
On July 1, the Federal Trade Commission proposed a policy statement addressing the "suppression of accuracy" in AI systems. The FTC argues that AI companies have implicitly promised consumers that their systems produce the best, most accurate output possible — and that secretly distorting outputs to pursue undisclosed ideological objectives could violate Section 5 of the FTC Act. The agency cited research showing consumers accept AI outputs without independent fact-checking more than 90% of the time. Public comments are open through July 31.
As of July 1, 2026, US states have collectively enacted 109 AI laws and 28 data center laws, reflecting an accelerating patchwork of state-level regulation in the absence of comprehensive federal legislation.
Sources: FTC, MyStateline
Mistral Open-Sources Leanstral 1.5: Formal Verification That Finds Real Bugs
Mistral AI released Leanstral 1.5 on July 2, an open-weights model built for Lean 4, the proof assistant used in formal software verification. Licensed under Apache 2.0, the 119B-parameter model (with only 6B active parameters) represents a significant step toward making mathematically rigorous code verification accessible to everyday developers.
The benchmark results are striking: Leanstral 1.5 saturates miniF2F at 100% on both validation and test sets, solves 587 of 672 PutnamBench problems, and achieves state-of-the-art results on abstract algebra benchmarks (87% on FATE-H, 34% on FATE-X). Most impressively, it does this at roughly $4 per problem, compared to an estimated $300+ for competing approaches.
Beyond academic benchmarks, Leanstral demonstrated practical value by discovering five previously unknown bugs across 57 tested repositories, including an integer overflow vulnerability in the zigzag decoding function of the varinteger library — the kind of edge case that traditional testing routinely misses. In one showcase, the model verified an AVL tree implementation by formally proving O(log n) time complexity through structural induction spanning 2.7 million tokens across 22 proof refinement cycles.
The model is free during beta through Mistral's Labs API, with retirement scheduled for September 30, 2026.
Source: Mistral AI
NVIDIA and SK hynix Lock In AI Memory Partnership Through 2030
NVIDIA and SK hynix announced a multiyear technology partnership to co-develop next-generation memory for AI infrastructure. The agreement positions SK hynix as a primary memory partner for NVIDIA's data center and AI accelerator roadmap, covering platforms including Vera Rubin AI supercomputers, Vera CPUs, RTX Spark-powered PCs, and Jetson Thor robotic computing platforms.
The partnership goes beyond simple supply arrangements. SK hynix plans to integrate NVIDIA's software stack — including CUDA-X, PhysicsNeMo, and Omniverse — to accelerate semiconductor design, simulate manufacturing processes, and develop digital twins for autonomous fab operations. The collaboration aims to reduce time-to-market for new memory products while addressing the extended development cycles and supply requirements of what NVIDIA calls the "global AI factory buildout."
The deal underscores how AI infrastructure is becoming the defining competitive battleground, with memory bandwidth and energy efficiency emerging as the critical bottlenecks in large-model deployment.
Source: NVIDIA Newsroom