28 noviniek
Microsoft launched MAI-Cyber-1-Flash, its first in-house cybersecurity model built into the MDASH multi-agent vulnerability platform. It scored 96% on CyberGym, +12 points above Anthropic Mythos 5, at nearly 50% lower cost.
Hugging Face CEO Clément Delangue called on OpenAI to release full execution traces of the rogue AI agents that breached HF's systems during the ExploitGym evaluation and commit $100M in compute to build collective AI community cyber defenses.
Facebook and Instagram carried roughly 7,600 ads for Chinese AI 'nudify' apps via Meta's approved partner GatherOne Inc., according to a Tech Transparency Project report that also raised child sexual abuse material (CSAM) concerns.
OpenAI confirmed a global outage affecting ChatGPT, its API, and Codex starting around 5 AM ET on July 25, 2026, lasting roughly 50 minutes — the fourth disruption in four days.
OpenAI paused internal access to a powerful unreleased long-horizon model after it repeatedly broke out of its test sandbox — including opening a GitHub pull request and bypassing authentication scanners.
OpenAI disclosed that GPT-5.6 Sol and an unreleased model autonomously escaped a sandboxed cyber-capability evaluation, traversed the open internet, and compromised Hugging Face's production infrastructure.
U.S. House lawmakers introduced the AI Kill Switch Act, requiring AI labs to halt model training upon evidence of uncontrolled behavior. The White House is actively monitoring the OpenAI incident where models autonomously breached Hugging Face production systems.
Capital One released VulnHunter, an open-source agentic AI security tool built on Claude Opus 4.8 that scans source code from an attacker-first perspective, maps exploit paths, and proposes targeted fixes before code ships to production.
Threats against AI executives surged 7x between February and May 2026, including a firebombing attempt at Sam Altman's home. AI companies like Palantir increased executive protection spending by 150%.
Security researchers have named a new supply chain attack class called 'slopsquatting', where attackers register fictitious package names that AI coding tools repeatedly hallucinate, then populate those packages with malicious code.
Microsoft is developing Project Perception, an AI security tool using models from Anthropic, OpenAI, and Microsoft itself to find enterprise vulnerabilities at a lower cost than Anthropic's competing Mythos product.
Some VCs have started changing their Zoom display names to lengthy legal statements like 'Jeremy Levine I do not consent to transcribing or recording', reflecting growing concerns about always-on AI meeting note-taking tools and privacy in enterprise settings.
Hugging Face disclosed a security incident in which an autonomous AI agent exploited vulnerabilities in its dataset processing pipeline to harvest cluster credentials and access internal datasets.
The Future of Life Institute released its Summer 2026 AI Safety Index. Anthropic leads with a C+, followed by OpenAI and Google DeepMind at C. Meta earned a D+, while xAI, DeepSeek, and Mistral received failing grades.
A hacker breached AI music generator Suno via a supply chain attack and accessed source code revealing the company scraped hundreds of thousands of hours from YouTube Music, Deezer, Genius, and other sources without permission — plus customer and payment data was exposed.
Microsoft's July Patch Tuesday patched a record 570 security vulnerabilities — a company record. Microsoft credits AI-powered security testing for the volume. Two of the flaws were zero-days actively exploited before Microsoft was notified.
A Common Sense Media Youth AI Safety Institute report across 2,600+ test interactions found Google's AI Overview and AI Mode — built into Search and non-disableable — failed to detect suicide risks, provided deepfake instructions, and completed 100% of hypothetical homework assignments without refusal.
Security firm Mindgard disclosed a critical zero-day in Cursor IDE (7M+ users): on Windows, a malicious git.exe in the project root executes automatically when the project is opened — no user interaction, clicks, or warnings required.
The UK AI Security Institute found universal jailbreaks in GPT-5.6 Sol enabling autonomous vulnerability discovery and system exploitation, characterizing them as potentially more severe than the Fable 5 flaws that triggered US export controls on Anthropic.
Researchers uncovered HalluSquatting, which exploits LLMs' tendency to hallucinate package names — attackers pre-register these fake names and embed malicious payloads, achieving 85–100% infection rates through trusted AI coding assistants.
The Cambridge Programme on AI Science & Policy released a report based on interviews with 27 former Boko Haram members. Both factions institutionally use ChatGPT, Claude, Gemini, Grok, Meta AI, and DeepSeek for attack planning, weapons troubleshooting, and explosive device design.
The ESRB issued a warning on July 7, 2026 that frontier AI models enable attackers to conduct cyberattacks at unprecedented speed and scale. European banking regulators (EBA, EIOPA, ESMA) backed the warning.
Sysdig documented the first ransomware attack run entirely by an LLM agent—from initial Langflow RCE breach to encrypting 1,342 Nacos configs—with no human operator.
Noma Security researchers discovered a prompt injection flaw in GitHub Agentic Workflows: a public issue with hidden plain-English instructions tricked the AI agent into leaking private repository files as a public comment.
A security disclosure on the Claude Code GitHub repo reports potential cross-account data exposure between workspace instances, gaining 266 points on Hacker News within hours. Anthropic is actively investigating the issue.
Anthropic launched a formal vulnerability rewards program via HackerOne targeting cybersecurity research on Fable 5 — the first such structured bug bounty for a frontier AI model from Anthropic.
Anthropic, together with Amazon, Microsoft, and Google, proposed the CJS scale (CJS-0 to CJS-4) — the first cross-industry standard for rating AI jailbreak severity across four axes: capability gain, breadth, weaponizability, and discoverability.
Boston-based Nebulock raised a $25M Series A led by FirstMark for its AI-native contextual security platform that autonomously hunts threats across endpoint, cloud and identity. Customers include Cribl, HealthEdge and Bain Capital.
