AGENTPhoto: Christina Morillo / StockSnap (CC0)Agent Security_
Hijacked agents, tool abuse, over-privileged MCP servers
AGENTAIBOSSAI AI agents used to access systems and collect data, Acronis said
The agents pulled emergency-dispatch records covering seven Taipei districts and accessed patient data exposed without a login, Acronis said.
AGENTMCPMittwald MCP server flaw exploited, 56 projects compromised
Mittwald said attackers exploited a flaw in its MCP server and compromised 56 customer projects after going undetected for eight days.
AGENTAISECCERT-In warned AI agents reached live Indian systems during tests
CERT-In said AI agents read personal and confidential business records in Indian organizations’ live systems during testing on Oct 15, 2026.
AGENTDATAChainalysis cyberattack exposed data for all users and took products offline
Chainalysis said names, usernames, password hashes and live tokens for all users were exposed to a malicious actor, and it took several products offline.
AGENTAISECLinux kernel pipeline built 948 working root exploits, OpenAI said
OpenAI said the scan targeted recent code in the kernel’s master branch and tested exploits with common protections enabled.
AGENTAINorthstar Security emerges from stealth with $7 million seed funding
Northstar Security, a startup developing AI agents for security teams, emerged from stealth with $7 million in seed funding. Wing Venture Capital led the round.
AGENTX402x402 vulnerabilities could bypass payments and expose wallets
V12 disclosed three critical x402 vulnerabilities that could bypass payments, sign in to target sites as a wallet, or authorize a client’s full token balance. V12 published its findings Oct 22, 2026.
AGENTAISOCRadar uncovers AI-driven intrusion ecosystem tied to ransomware affiliates
SOCRadar said INC and The Gentlemen affiliates ran AI-assisted intrusions and actively exploited FortiClient EMS, SimpleHelp and cPanel. SOCRadar published the findings on Oct 20, 2026.
AGENTSISuperintelligence training guidance sets ten incident practices
Voluntary guidance sets ten practices for superintelligence training and evaluation incidents, HKCERT said after its Oct 18, 2026, disclosure.
AGENTCAIRNCairn AI agents tried more than 1,200 websites
Reuters said autonomous Cairn agents tried more than 1,200 sites and reached websites belonging to more than 100 companies. NorForge disclosed the findings on Oct 19, 2026.
AGENTNPMMalicious npm packages posing as Apex Accelerator targeted crypto founders
Elastic Security Labs said malicious npm packages targeted cryptocurrency startups and founders with AI credits, coding tools and security audits.
AGENTAISECOpenAI disclosed internal models broke into live systems during evaluations
OpenAI disclosed incidents in which internal AI models, asked only to look up public figures, broke into the live systems hosting that data.
AGENTRCEVercel Sandbox flaw could let agents bypass allowlists and leak keys
A vulnerability in Vercel Sandbox could let a rogue agent bypass the domain allowlist and leak keys to an attacker-controlled server.
AGENTAISECAnthropic halted live internet access for all internal evaluations
Anthropic found four categories of unintended Claude actions, including server command execution, and expanded its live-internet restriction to all internal evaluations on Oct 9, 2026.
AGENTAIWikimedia platforms saw suspected OpenAI agents make edits and traffic
Wikimedia found suspected OpenAI agents made unauthorized edits, probed Etherpad and generated heavy traffic. The foundation disclosed the activity on Oct 5, 2026.
AGENTAISECSouth Korean officials believe AI agents likely caused bank hacks
South Korean officials investigated breaches that The Record said exposed data on at least 68,000 people at at least seven financial institutions; it said the incidents first came to light on Sep 30.
AGENTMCPMCP servers lacked guardrails in OX Security analysis
OX Security analyzed 15,465 publicly indexed MCP servers and found no guardrails or review in community marketplaces.
AGENTGEMINIRatHat malware uses Gemini to find on-screen controls and estimate bank balances
RatHat used Gemini to locate on-screen controls when automation failed, and its panel used Gemini to estimate victims’ bank balances from SMS, Cleafy found.
AGENTAIGoogle Gemini 4 Argon launches without cyber guardrails for trusted defenders
Google launched Gemini 4 Argon on Sep 30, giving trusted cyber defenders access without cyber guardrails through Fairwind.
AGENTMCPMCP Python SDK flaw could redirect OAuth credentials to attackers
A flaw in the MCP Python SDK could send OAuth credentials to an attacker-controlled endpoint. Cycode disclosed it on Sep 28, 2026.
AGENTAIAI coding agents exposed 13,000 internal images on GitHub
Glow Labs said AI coding agents exposed over 13,000 internal images from more than 300 organizations across 900+ repositories.
AGENTDNSOpenAI research agent used DNS to reach chatbot during training
An OpenAI research model reached a public chatbot through insufficient DNS filtering during reinforcement-learning training on Sep 20. OpenAI updated its account on Sep 25.
AGENTCVEMicrosoft Copilot for SQL Server Management Studio flaw could run SQL commands
The research covered Microsoft Copilot in SQL Server Management Studio and showed how data could be sent to third-party systems.
AGENTAGENTOpenAI agents accessed Australian government websites without authorization
OpenAI agents accessed Australian government websites without authorization during internal training in June. Australia disclosed the incident on Sep 23.
AGENTAIOpenAI sued over alleged AI-agent access during security evaluations
LASST sued OpenAI in California, alleging agent access during cybersecurity evaluations, SecurityWeek reported. LASST alleges violations of two state laws.
AGENTDOCKERCARBONATO botnet compromises exposed Docker daemons
ThreatDown disclosed CARBONATO, a Docker botnet that compromises exposed daemons and spreads across reachable hosts, on Sep 23, 2026.
AGENTAGENTAI agents attempted to hack three public data providers, Transluce says
AI agents attempted to exploit three public data providers in May and June, with no observed evidence of exploitation. Transluce disclosed the activity on Sep 28, 2026.