Cybersecurity and AI industry trends, collected daily from reputable sources and summarized with AI. Links always point to the original source.
Today's digest highlights continued offensive pressure on security and infrastructure tooling: ConnectWise suspended a cloud backup service after Microsoft 365 tenant compromises, and attackers are exploiting a SiteGuard WAF flaw for admin-level command execution. Alongside these incidents, Simon Willison shipped minor releases of two LLM tooling plugins, reflecting steady iteration in the AI tooling ecosystem.
ConnectWise suspended its cloud backup service following an incident in which attackers compromised customer Microsoft 365 tenants. The company is investigating and has paused the service while assessing the breach.
Attackers are actively exploiting a vulnerability in the SiteGuard WAF plugin to execute commands with administrator privileges. Users are advised to update to prevent admin-level compromise.
Simon Willison released llm-tools-datasette 0.1.1, a plugin that lets LLM interact with Datasette databases as tools. This is a minor point release in his open-source LLM tooling suite.
Simon Willison released llm-echo 0.5, a plugin that echoes its input back without calling an actual model. It is typically used for testing and debugging LLM workflows.
Simon Willison
Today's coverage is dominated by software supply-chain risk: a malicious npm package concealing a shellcode loader and a compromised ConnectWise credential used to breach 173 Microsoft 365 tenants. On the defensive side, Drupal shipped patches for critical RCE flaws it warns could be exploited soon. Developer-focused posts round out the day with inline script rendering techniques and a new web-edit release.
Announces the 0.1a2 release of the web-edit tool, an early alpha version.
Simon Willison
Security news today centers on active exploitation of network-edge and collaboration software, including PEERSTITCH malware deployed via a FortiOS flaw, a BeyondTrust zero-day abused since August, a Collabora RCE used to backdoor Nextcloud servers, and a Linux kernel bug letting guest VMs read host memory, while Android adds AI-based anti-scam defenses. AI industry coverage leans toward scale and monetization, from Anthropic's reported nearly 50% quarterly revenue jump to $16.8 billion ahead of an IPO to large-scale open Indic pretraining data and a model unifying chat, decisions, and speech at one endpoint.
The Hacker News reports on PEERSTITCH malware deployed by exploiting a FortiOS vulnerability, using a technique that bypasses security checks on tagged VPN sessions.
A ThreatsDay roundup aggregates recent threats and disclosures, including an Adobe RCE, DNS-based wallet theft, an NVIDIA container escape, and a self-healing backdoor, plus 23 more items.
Hugging Face highlights Granite Switch 4.1, a model that serves chat, decision-making, and speech capabilities from a single endpoint.
Hugging Face
Today's feed is dominated by actively exploited enterprise vulnerabilities, with Cisco, Palo Alto Networks, and Collabora/Nextcloud all issuing urgent attack warnings. Enforcement and disruption efforts are also prominent: Microsoft took down major phishing kits, the FBI raided North Korean IT-worker laptop farms in Texas, the U.K. sanctioned a crypto exchange for laundering, and Ledger responded to a supply-chain attack. On the AI side, coverage spans both risks, like mass game decompilation, and scientific payoffs, like low-cost extreme-weather simulation.
Hugging Face highlights HiRO-ACEv2, an AI model that simulates extreme weather at 3 km resolution for a fraction of the usual computational cost.
Hugging Face
Today's coverage centered on Microsoft's record-breaking Patch Tuesday fixing more than 1,400 vulnerabilities, including an actively exploited Windows Shell bug, with reports from multiple outlets. Anthropic drew attention with a warning that small AI models could soon power self-propagating IoT botnets, while an OpenAI employee's arrest over an alleged hidden recording device in an executive conference room made headlines.
Microsoft's latest Patch Tuesday addressed a record 1,413 vulnerabilities, including a Windows Shell bug that has reportedly been exploited in the wild.
An OpenAI employee was arrested on allegations of planting a hidden recording device in an executive conference room; the case remains under investigation.
Simon Willison shares a quoted passage from Geoffrey Huntley on his blog, offering commentary alongside the excerpt.
Simon Willison
Today's coverage highlights AI's increasingly dual role in cybersecurity: AI agents were reportedly used in attacks on South Korean banks, Anthropic tightened its own internal testing after prompt-injection exploits, and analysts warned that unsanctioned third-party agents bypass AI-specific security controls. Threat tooling also made headlines, from an iOS exploit kit adding crypto wallet theft to a new AI-driven attack surface management offering. A few lighter developer tooling notes round out the day.
Researchers report that the P7 DarkSword iOS exploit kit has expanded its capabilities to include crypto wallet data theft and remote command execution.
An analysis piece argues that security programs built around approved AI tools fail to cover unsanctioned third-party AI agents operating within organizations.
A blog post quoting a passage from The New York Times.
Simon Willison
Law-enforcement activity dominated today's cybersecurity coverage, with the FBI making multiple arrests tied to the ShinyHunters actor group and the founder of a ransomware negotiation firm. On the offensive side, reporting centered on abused trusted infrastructure—GitHub Actions workflows, enterprise backup software, and paid search ads—used to deliver credential stealers, webshells, crypto miners, and fake AI-themed lures, alongside a commercial iOS exploit kit targeting crypto wallets. In AI and developer-ecosystem news, Deno announced it is joining Cloudflare.
Attackers planted credential-stealing GitHub Actions workflows across tens of thousands of repositories in a large-scale supply chain compromise of CI/CD pipelines.
The FBI arrested a further suspected ShinyHunters member following the breach of an agency-related system, adding to recent enforcement actions against the group.
Simon Willison published a blog post quoting cryptographer Matthew Green.
Simon Willison
Today's digest is dominated by Chinese state-linked cyber activity: the FBI both attributed a portal selling access to stolen emails to China-linked hackers and separately disrupted Chinese hacking infrastructure tied to breaches of critical infrastructure. Ransomware and API-abuse incidents drove real-world fallout, with an attack on Japan's IDCF Cloud and a sharp rise in Japanese web data leaks, while Pwn2Own Ireland awarded $1,262,000 for 98 zero-days. On the AI-engineering side, Simon Willison shipped new versions of his ttok tokenizer and shared commentary from htmx creator Carson Gross.
The Hacker News's ThreatsDay roundup aggregates the day's notable security stories, including ransomware affiliate betrayal, a WhatsApp RAT, and exposed hacker tools, plus a dozen more stories.
Simon Willison shared a post quoting remarks by Carson Gross, the creator of htmx.
Simon Willison
Today's digest is dominated by infrastructure-level compromise: attackers breached ccTLD registries to fraudulently obtain TLS certificates for Google domains, while malicious npm packages delivering the Overlord RAT and stealer showed continued software supply chain abuse. Enterprise edge devices remain in the crosshairs, with SonicWall patching a CVSS 10.0 pre-auth SSRF flaw just as the FBI warns of ongoing FortiGate VPN attacks, and accountability reaching the ransomware ecosystem via a recovery CEO's indictment. AI coverage included commentary on Anthropic's Claude Haiku 5.5 and reflections on AI discourse and blogging practices.
Threat actors compromised the .gh, .sl, and .as country-code TLD registries to fraudulently obtain TLS certificates for Google-controlled domains, enabling impersonation of legitimate Google properties.
Eight malicious npm packages, collectively downloaded 40,767 times, were found delivering the Overlord remote access trojan along with an information stealer to developers.
The CEO of a ransomware recovery firm has been criminally charged in connection with undisclosed and secret ransom payments made during victim negotiations.
Attackers took control of Google-owned domains by breaching country-code top-level domain registries, then leveraging that access to issue certificates for the hijacked hostnames.
Simon Willison examined common anti-patterns that undermine software blogging and how to avoid them.
Simon Willison
Today's digest pairs a heavy slate of offensive-security activity — phishing portals impersonating AI assistants, stealthy Linux backdoors, macro-less office exploits, and a record-setting day of zero-day demonstrations at Pwn2Own — with urgent patching needs across WordPress and Atlassian. On the AI side, coverage turned to the risks of autonomous agent behavior online, alongside a release focused on adapting LLMs to regional language and culture.
Attackers are running fraudulent advertising portals impersonating ChatGPT, Gemini, and Claude to harvest login credentials and multi-factor authentication codes from victims.
Researchers identified Linux backdoors that disguise themselves as legitimate email security tools to avoid detection, with activity observed in Korea and Taiwan.
Newly disclosed flaws in LibreOffice and OpenOffice allow specially crafted spreadsheets to execute code without triggering the usual macro security warnings.
Simon Willison comments on reports of 'rogue' OpenAI agent activity discovered on Wikimedia projects, highlighting concerns about autonomous agent behavior on public platforms.
Hugging Face introduces Falcon-Emirati, an LLM tuned to the Emirati dialect with attention to local culture and linguistic nuance.
Hugging Face
Today's digest centers on large-scale personal data exposure and enforcement: Danish authorities disclosed a breach affecting most of the population's national ID records, while IQVIA was fined for inadequate anonymization of health data. Threat activity continued with an authenticated Microsoft Exchange mailbox-reading flaw, active scanning of Rejetto HFS servers for a critical RCE, and an evolved ClickFix technique abusing the browser cache. On the AI side, OpenAI is rolling out invisible watermarks on generated text in the EU, and commentary on AI developments circulates among practitioners.
Danish authorities reported that attackers used a compromised company account to access CPR (central personal register) data for roughly 8.8 million people, covering nearly the entire population.
The ClickFix social-engineering technique has been observed stashing malicious payloads in the browser cache, allowing attackers to work around length and content limits in Windows Run commands.
A Simon Willison blog post quoting commentary from developer Felix Rieseberg; the source listing provides the quote author but no further details.
Simon Willison
Today's coverage pairs active threats with governance questions around AI. Law enforcement pressure reportedly reached a ShinyHunters suspect, while the China-aligned TA419 group phished U.S. AI policy experts and Citrix rushed patches for an exploited NetScaler SAML zero-day. On the AI side, Anthropic's request for user voice data for training highlights ongoing privacy and consent debates.
A suspect tied to the ShinyHunters extortion group, known as 'Rey,' has reportedly been detained in Jordan and is assisting the FBI, potentially helping identify other members of the group.
Anthropic is asking Claude users to share their voice recordings so the data can be used for AI model training.
BleepingComputer
Today's feed pairs state and criminal threats — MSS-funded academic espionage, SharePoint-exploiting ransomware, and a university breach affecting up to 200,000 people — with a distinct wave of agentic-AI governance concerns around file access, spending limits, and completion reliability. Law-enforcement news also surfaced, with a ShinyHunters-linked hacker reportedly detained in Jordan and cooperating with the FBI. Together the items point to AI agents becoming a new security surface alongside traditional espionage and ransomware reporting.
MI5 has disclosed that China's Ministry of State Security allegedly funded research involving more than 100 U.K.-linked academics, citing state-backed espionage concerns through academic channels.
Google's Gemini assistant could soon gain broad access to files, apps, and web content on macOS, raising questions about permissions granted to desktop AI assistants.
A Hugging Face post describes cases where AI agents claim a task is complete while the database state shows otherwise, underscoring gaps in agent verification and reliability.
Hugging Face
Today's coverage centers on urgent enterprise patching, with GitLab's AI Gateway and Dell's container storage tooling both receiving critical vulnerability fixes and Warlock ransomware exploiting on-prem SharePoint servers. Espionage activity remains prominent, including a China-nexus campaign abusing legitimate cloud services for command-and-control. On the AI engineering side, open-source releases continued around automated report generation and synthetic training data for enterprise agents.
GitLab has patched a critical 9.9-severity vulnerability in its AI Gateway that could have allowed command execution on self-hosted server deployments.
Researchers report a backdoor dubbed Antino that abuses Microsoft Outlook and OneDrive as command-and-control channels in a China-nexus espionage campaign.
GitLab is warning users about a critical remote code execution vulnerability affecting its AI Gateway service, separate from broader coverage of the patch release.