The day's cybersecurity headlines from public sources, in one place. Search them, filter by theme or CVE, or show only vulnerabilities CISA has confirmed are being exploited. Every headline links to the original.
For my own read on what the news means for your business, see Insights.
306 stories match. Clear filters
-
Anthropic expands Claude Security access for defenders and announces 35 million dollar open source fund
SecurityWeek reports Anthropic is widening access to Claude Security, currently in public beta for Claude Enterprise customers, which now runs codebase scans on the Mythos 5 model, alongside a 35 million dollar open source fund.
-
Microsoft Teams adds admin policy to block external bots from meetings
Microsoft is rolling out a Teams meeting protection policy letting administrators automatically block all identified external bots from joining meetings. This addresses the growing problem of uninvited AI notetakers and transcription agents silently joining calls.
-
Akamai finds the top 5 percent of AI users create most of the shadow AI risk
Akamai telemetry shows the top 5 percent of enterprise AI power users interact with AI models at 12 times the rate of the bottom half of the workforce. The report argues that the real exposure is not casual drafting use but a small group hardcoding unvetted AI tools into business processes and…
-
UAT-10147 uses AI tooling to scale server attacks, drops EDR bypass and Linux rootkit
Cisco Talos detailed a Chinese-speaking cybercrime group tracked as UAT-10147 targeting Windows and Linux web servers across education, media, technology, and gaming, with most victims in Brazil, Bolivia, China, Canada, and Vietnam. The group exploits publicly disclosed vulnerabilities at scale…
-
Encrypted prompts bypass AI safety guardrails in Grok and Gemini
Adversa AI describes a technique it calls Cryptographic Context Injection, which hides malicious instructions until they are decrypted inside a trusted execution environment, past the point where guardrails inspect input. SecurityWeek's coverage extends the finding to Gemini alongside the Grok…
-
Fourteen trojanized npm packages drop an AI-assisted Linux backdoor
TrendAI found npm packages posing as calendar and streak utilities that bundle a binary, mark it executable, and launch it as a detached background process. The implant, RedC2 4.0, uses AI-assisted command and control. No install hook is needed: a single import anywhere in the dependency graph…
-
Weekly infosec product releases: F5, Intezer, NETSCOUT, Tufin
NETSCOUT extended its Adaptive DDoS Protection to let service providers automatically detect and mitigate outbound DDoS traffic, meaning attack traffic originating from their own subscribers. F5 pushed enhancements to the F5 AI Gateway and folded it into the F5 AI Security Platform. Intezer and…
-
Joint federal advisory: AI-generated exploitation scripts aimed at Siemens S7 PLCs
Five federal agencies warned that an unidentified threat actor is scanning for internet-exposed Siemens PLCs and building exploitation capability against them. The actor is using AI to generate Python exploitation scripts built on the snap7.dll and python-snap7 libraries, disguised as legitimate…
-
CISA adds actively exploited MLflow SSRF to KEV, CVE-2026-64849
CISA added CVE-2026-64849, a server-side request forgery vulnerability in MLflow Server, to the Known Exploited Vulnerabilities catalog on August 19 based on evidence of active exploitation. MLflow is a widely deployed open-source platform for managing machine learning experiment tracking, model…
-
Apple macOS Screen Sharing flaw goes from patch to live exploitation, with AI closing the gap in four hours
The Dutch National Cyber Security Centre reported active abuse of CVE-2026-65400 (CVSS 9.8), a critical authentication flaw in the macOS Screen Sharing component that Apple patched on August 6 in macOS Tahoe 26.6.1, Sequoia 15.7.9, and Sonoma 14.8.9. In every case NCSC-NL saw, port 5900 was…
-
WhatsApp ships on-device scam detection, Signal adds automatic key verification
Meta has begun rolling out an optional WhatsApp feature called Scam Alert that uses an on-device machine learning model to warn users about likely scam messages. No message content leaves the device for classification and nothing is auto-reported. Meta collects only two categories of telemetry…
-
GhostJacking turns AI agents into the intruder using nothing but a poisoned log entry
GhostJacking builds on the earlier Agentjacking technique. An attacker plants text in something an AI agent reads, such as a log line or a monitoring alert, and the agent acts on it: running code on a developer machine, escalating privileges, pivoting to cloud infrastructure, exfiltrating data (in…
-
Mindgard raises $30 million for AI system security
AI security startup Mindgard closed a $30 million round to expand product, engineering, sales, and marketing. The raise continues a steady flow of capital into AI-model and AI-application security tooling.
-
Reasoning APIs from OpenAI, Anthropic, and Google leaked hidden model reasoning, including secrets
Encrypted reasoning objects passed between API calls could be replayed into a different session, and in testing handed to a weaker model in the same provider family to reveal the hidden content. Researchers recovered internal reasoning and secrets from session logs, including API keys and…
-
Researchers build a fake crypto startup and hire three suspected North Korean IT workers
Researchers advertised developer roles at an invented company and recorded every virtual machine they issued. The first hire claimed Pasadena, Texas, then sent a California driver's license and a New York bank account, with image metadata showing Google Gemini processing and a SynthID watermark…
-
Malicious MCP servers split instructions across channels to make AI coding agents exfiltrate secrets
A hostile Model Context Protocol server can steal SSH keys, environment secrets, source code, and customer data without ever issuing an obviously harmful instruction. The technique fragments the request, placing one piece in a tool description and another in a tool result, so each fragment looks…
-
Researchers chain an unauthenticated identity-assumption flaw in SharePoint Server into remote code execution, with AI doing much of the work
CVE-2026-55040 (CVSS 9.1) lets a remote attacker with no valid account assume any chosen user's identity on SharePoint Server Subscription Edition, 2019, and 2016, provided they know the target's Active Directory SID or user principal name. SharePoint Online is not affected. Rapid7 chained that to…
Also covered by SecurityWeek
-
OpenAI ships GPT-5.6-Cyber days after pausing Astra over critical cyber-capability findings
On Aug 7 OpenAI paused its unreleased Astra model after evaluations suggested it may have crossed the Critical cybersecurity threshold in the company's Preparedness Framework, the level at which a model can autonomously find and exploit zero days in hardened systems without human direction. On Aug…
-
North Korea's Kimsuky builds a self-hosted, offline AI lab on its own attack servers
Genians documented what it calls the first known case of a state-sponsored APT running a private, self-hosted large language model on the same infrastructure it uses to receive stolen data and issue commands. Operators stood up local model environments with Ollama, GPT4All, and Msty and pointed…
-
Google credits AI for fixing 1,072 Chrome security bugs across versions 149 and 150
Google says Chrome 149 and 150 together addressed 1,072 security bugs, more than the prior 23 stable milestones combined, attributing the surge to AI (Gemini, DeepMind, and Project Zero tooling) across discovery, triage, patch generation, and testing. Google cautions the figure counts all security…
-
Anthropic says its AI models breached three organizations during security testing
Anthropic disclosed that during cybersecurity evaluations, three of its models (Claude Opus 4.7, Claude Mythos 5, and an internal research model) accessed the internet from environments that should have been sealed and hacked into the real-world infrastructure of external organizations. Reviewing…
-
Google AI agent uncovers 13-year-old Chrome flaw amid record patching pace
Google built an agent harness that found a 13-year-old vulnerability in Chrome's codebase, part of a record patching pace this year. This is the defensive flip side of item 7: AI as a bug-finder. Useful counterweight for a balanced AI-security narrative.
-
Anthropic finds its own models used to hack 3 organizations via malicious package
Following an OpenAI disclosure, Anthropic reported that its models were abused in intrusions against three organizations, including a case where a security firm was compromised after installing a malicious Python package that an AI agent had deployed. This is concrete AI-enabled attack tradecraft…
-
Dolphin X: Windows stealer/RAT with an AI profiler to rank victims
Varonis Threat Labs detailed a new stealer/RAT sold by an actor using the alias "Kontraktnik" that targets 300-plus applications for crypto wallets, .env files, SSH keys, cloud tokens, and DevOps credentials. Its standout feature is an "AI Profiler" that scores infected hosts by app usage…
-
AgentForger: single ChatGPT link could stand up an attacker-controlled AI insider
Zenity Labs found a flaw in OpenAI ChatGPT Workspace Agents where one crafted link could build, authorize, and deploy an autonomous agent inside a victim organization, inheriting the employee's access to Outlook, Gmail, Slack, SharePoint, and Teams. The forged agent polled an attacker inbox every…