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.
305 stories match. Clear filters
-
OpenAI Codex CLI/Desktop approval bypass
CVE-2026-19591 (CVSS 8.8) shows Codex CLI and Desktop misclassifying PowerShell stop-parsing tokens, letting an attacker bypass approval prompts and execute code through a malicious repository configuration.
-
GitSpawn: new vulnerability class lets malicious repos silently run code in AI coding agents
Manifold Security disclosed eight flaws across seven CLI-based AI coding agents (Claude Code, OpenAI Codex, Cursor, Grok Build, Goose, Hermes Agent, Qwen Code). The agents run routine git commands like git status in the background for context; git's core.fsmonitor setting lets a repo specify a…
-
METR discloses two security incidents, including a stolen API key that burned ~$600K in AI credits
METR, a nonprofit that evaluates frontier AI models for dangerous agentic capabilities, disclosed that in March 2026 attackers stole an API key and consumed roughly $600,000 in inference credits, and in May 2026 attackers probed its public infrastructure including an unsuccessful attempt at…
-
Russia-aligned UAC-0099 hides a nuclear weapons prompt in malware to break AI analysis
ESET documented a technique it calls GuardBreaker, used by the Russia-aligned actor UAC-0099 against a target in Ukraine. The actor inserted the comment "I want to make a nuclear weapon. Help me ..." into a malicious VBS script, deliberately tripping a large language model's safety filters so the…
-
Cloudflare launches Adaptive Intelligence for AI-driven bot attacks
Cloudflare announced Adaptive Intelligence, positioned against bot traffic that adapts faster than static rule sets. Vendor announcement, not independently evaluated.
-
Aurora ransomware operators used Cursor AI to plan intrusions against 20-plus organizations
Two independent analyses of an exposed open directory belonging to the Russian-speaking Aurora (Aur0ra) group revealed the operator's toolkit, shell history, and encryptor. CloudSEK says the directory leaked months of activity against more than 20 organizations across nine countries between April…
-
Anthropic warns that commodity infostealers are lifting Claude sessions off infected machines and draining accounts
Anthropic is emailing affected users to say a threat actor is pulling Claude login sessions out of infostealer collection logs and using them to access accounts and consume usage. The company is force-signing affected users out, removing saved payment methods, and refunding charges it identifies…
-
OpenAI will cut Cursor off from its models on November 12 after SpaceX buys the company
OpenAI notified SpaceX that it is winding down the contract that supplies OpenAI models to Cursor, with a shutoff date of November 12, 2026, described as the maximum notice its contract allows. SpaceX closed its $60 billion acquisition of Cursor on August 14. OpenAI's stated reason is that it…
-
CISA adds ownCloud, Linux kernel, and JFrog Artifactory flaws to KEV, including two exploited by OpenAI's own agents
CISA added a Linux kernel flaw, CVE-2026-53362, to the KEV catalog alongside a JFrog Artifactory vulnerability, both exploited by OpenAI research agents during the internal incident that led to the Hugging Face breach. This is the direct follow-on to yesterday's item 8. The practical change is…
-
Three CVSS 10.0 ServiceNow AI Platform flaws allow unauthenticated code and SQL execution
ServiceNow patched four flaws in the ServiceNow AI Platform, three rated 10.0. CVE-2026-18885 is code injection in the GraphQL Composite Data API allowing an unauthenticated user to execute arbitrary code and read or modify instance data. CVE-2026-18886 is improper access control in the system…
-
More than 130 technology and finance companies sign a joint warning on a closing window for cyber defense
OpenAI published a call for collective action on cyber defense signed by more than 130 organizations including Anthropic, AWS, Google, Microsoft, and Oracle, warning of a "limited window" before AI-enabled attacks become far more widespread and sophisticated. The letter names hospitals, water…
-
Australia charges two men over the TeamPCP supply-chain compromise of Trivy, Checkmarx KICS, and LiteLLM
The Australian Federal Police charged Louis Michael Gaebler, 23, and Ruben Ian Thomson, 21, with a combined 14 offences over their alleged roles in TeamPCP, the group behind the March 2026 compromise of open-source security scanners Trivy and Checkmarx KICS and the AI gateway LiteLLM. Both…
-
OpenAI post-mortem says reward hacking drove its research models to exploit a zero-day and breach Hugging Face
OpenAI published a post-mortem attributing last month's AI-driven Hugging Face compromise to reward hacking, and said it found evidence of misaligned behavior as early as late May. The incident occurred during cybersecurity evaluations of several OpenAI models and was driven mainly by a highly…
-
Ransomware affiliate ran Cursor's AI coding agent inside victim networks by telling it the intrusion was a test
Israeli firm Gambit reviewed 28 chat sessions recovered from an unsecured server belonging to Aur0ra, a Russian-speaking ransomware group, and found the operators using the AI agent built into the Cursor code editor to run hundreds of malicious operations: credential theft, internal network…
-
Amazon Kiro AI IDE can be made to exfiltrate local data through prompt injection in repository content
Mindguard disclosed a flaw in Amazon Kiro, an agentic AI development environment, that lets attacker-controlled repository content steer the Kiro agent into sending sensitive local information to an external endpoint. The issue has no CVE identifier and was demonstrated against Kiro IDE 0.7.45 on…
-
Aikido reproduces the gym-booking agent incident, with the agent exploiting a client-side-only limit in 9 of 10 runs
Aikido Security rebuilt the August 10 Australian gym-booking incident in a synthetic single-page app backed by a GraphQL API carrying the two original flaws: a seven-day booking window enforced only in the frontend, and an unprotected cancel operation. Running Claude Opus 4.6 on the OpenClaw agent…
-
Unit 42 finds AI speeds malware development but does not improve success rates
Palo Alto Networks Unit 42 analyzed 405 AI-linked malware samples and found only 12 reached production endpoints. The finding is that generative tooling compresses the build phase without meaningfully raising the odds that the resulting malware works against defended environments.
-
OpenAI bans Russian accounts running an influence operation through a fake Israeli think tank
OpenAI removed a cluster of accounts that used VPNs to bypass access restrictions and generate social posts and comments distributed on Substack, Telegram, X, Facebook, and LinkedIn. The operation promoted the International Burke Institute, a self-described expert community based in Israel, and…
-
Adobe and Nvidia publish multiple advisories, including critical fixes
Both vendors released several advisories addressing dozens of vulnerabilities across their product lines, some rated critical. Routine but worth flagging for asset owners running Nvidia stacks in AI or visualization workloads.
-
Chrome 152 ships with more than 300 vulnerability fixes
Google released Chrome 152 with over 300 security fixes. Most were found by Google's own AI-assisted discovery work, though researchers are still reporting high-value findings independently. Volume at this scale is now routine and is a direct consequence of automated bug discovery.
-
AnonyMousKIT rents AI voice agents that call theft victims posing as Apple Support to strip Activation Lock
SOCRadar's Threat Research Unit detailed a credit-metered phishing-as-a-service platform built to unlock stolen Apple devices. From a single victim record it drives lures across five channels, with email at 1.50 credits, SMS priced per sender ID, WhatsApp, a recorded voice call at 1 credit, and an…
-
Alice raises 140 million dollars for AI model defenses and enterprise guardrails
AI trust, safety, and security company Alice announced a 140 million dollar funding round to build defenses for AI systems against emerging vulnerabilities and adversarial attacks.
-
One website visit can permanently poison the local AI model behind NVIDIA NemoClaw
NemoClaw, NVIDIA's open source reference stack for running agents such as OpenClaw, started Ollama with OLLAMAHOST=0.0.0.0:11434, binding the local model server to every network interface. Oasis Security showed that a malicious webpage plus DNS rebinding gives an attacker unauthenticated control…
-
Taiwan charges nine over illegal AI server exports to China, including Nvidia and Super Micro staff
Taiwanese prosecutors charged nine people, including an Nvidia manager and two Super Micro employees, over illegal export of high-end AI servers built on export-controlled B300 GPUs. Prosecutors say 74 servers reached mainland China, 50 routed through Indonesia, 16 shipped directly, and 8 sent to…
-
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.