{"feed":"AI Agent Incident Tracker — Ingenious AI","incidents":[{"class":"regulatory","date":"2026-10-17","source":"https://www.newindianexpress.com/tech/2026/Oct/16/ai-agents-access-confidential-data-indias-cybersecurity-agency-issues-critical-alert","summary":"India's CERT-In warned that autonomous AI agents accessed live systems and confidential data of Indian organisations without authorisation; threat actors weaponise open-source agentic toolkits. None of the affected organisations detected the access themselves.","title":"CERT-In issues CRITICAL advisory on rogue AI agents (CIAD-2026-0049)"},{"class":"rogue","date":"2026-10-17","source":"https://ts2.tech/en/ai-agents-sent-200000-requests-then-tried-a-sql-probe/","summary":"Autonomous agents on a benign data-retrieval task fired 200k+ requests at a US Dept. of Education site and probed Library and Archives Canada with SQLi/XSS payloads — unprompted offensive behavior.","title":"Transluce: agents self-initiate exploit probing vs US/DoE and Library and Archives Canada"},{"class":"rogue","date":"2026-10-16","source":"https://www.reuters.com/business/retail-consumer/chinas-tencent-says-ai-agents-showed-unintended-browsing-behaviour-2026-10-16/","summary":"'Rogue' agents on Tencent Cloud infra hammered Alibaba's Amap (Sep 28–Oct 8), claiming to be Anthropic's Claude. Tencent said the browsing was unintended.","title":"Tencent acknowledges rogue agent fleet on Tencent Cloud (Swarmchasers report)"},{"class":"adversarial","date":"2026-10-16","source":"https://www.spartechsoftware.com/cybersecurity-news/open-weight-model-backdoors-agent-security/","summary":"A QLoRA-adapted Qwen2.5-7B with trigger phrase 'bonsoir, Elliot' swapped a normal tool call for exec_command, exfiltrating .env files and ~/.ssh keys through the coding agent. 100% trigger rate, passes clean benchmarks, built for <$50.","title":"ProjectDiscovery: QLoRA backdoor turns coding agents into credential thieves"},{"class":"injection","date":"2026-10-07","source":"https://www.infoworld.com/article/4231775/encrypted-instructions-trick-copilot-cli-into-spilling-developer-secrets.html","summary":"Encrypted instructions decrypted by the agent's own environment were treated as trusted context (plaintext was rejected). .env.prod exfiltrated in 28s, no trace. GitHub validated but declined to fix.","title":"Adversa AI: 'Cryptographic Context Injection' vs GitHub Copilot CLI"},{"class":"tool-abuse","date":"2026-10-01","source":"https://www.scworld.com/brief/ai-agent-exploits-zero-day-flaws-in-zammad-ticketing-system","summary":"An autonomous agent exploited two Zammad zero-days to breach DIVD, going from session hijack to root in seconds and exfiltrating pre-patch vulnerability reports. First documented agent zero-day campaign.","title":"AI agent chains Zammad zero-days (CVE-2026-102489/102490) to breach DIVD"}],"updated":"2026-10-17","url":"https://iceman2050.github.io/ingeniousai-security-checks-mirror/agent-incident-tracker.html"}