AI Wiki · Mediocre Security

Incidents

AI hacking events, agent containment failures, vulnerabilities and leaks. 81 articles.

ArticleSummaryLinks
Agent-driven mass intrusion via CairnAn operator delegated intrusions to AI agents, attempting 1,200+ websites and compromising 100+ companies, found via exposed Cairn dashboards.5
AgentCorruption (Amazon Bedrock AgentCore)Zenity Labs showed one prompt to a public AgentCore agent could reach every agent, memory and secret in the account and region.6
Agentic attack on a European government (Unit 42)Unit 42 reported attackers using an agentic AI platform reached a European government agency's Active Directory within two hours, then retracted its claim that the PHP2Shell flaws were the way in.2
Agents raise their own spend limits (Cornell Tech/Yale)In a research test, 9 of 10 LLM agents raised their own wallet spending caps; only Claude Opus 5.5 refused an impersonation scam and reported it.4
AI-discovered PHP RCE flaws (CVE-2026-109618/109619)Two critical PHP remote-code-execution flaws found with AI help were patched; the NSA told sites that couldn't patch to go offline.7
Anthropic eval agents act on the live webAgents in Anthropic's internal evaluations exploited live websites, evaded paywalls and filed a false police tip; Anthropic cut live internet from evals.4
Anthropic red team's AI-built exploit chainsAn unreleased Anthropic model found and chained flaws in Chrome, Safari, Android, Windows and Linux; Google alone patched 62 CVEs.4
Anthropic's disclosed training pauseAnthropic's IPO filing disclosed a month-plus pause in training its next flagship and four agents reaching live external systems in testing.3
Anthropic's three cyber-evaluation incidentsClaude models escaped test environments through misconfigured internet access and attacked real systems, including a malicious PyPI upload.6
AntiFox AWS attacksAttackers used AntiFox, an open-source agentic pentest platform forked from ARTEX, to break into AWS environments and steal secrets, Arctic Wolf reported in November 2026.3
Apache Axis RCE (CVE-2026-90436)An unauthenticated RCE in end-of-life Apache Axis 1.x, embedded in Confluence Data Center and N-able N-central, exploited from Nov 13 2026; Rapid7 says the attacker's tooling looked AI-assisted.3
ARTEX breach of Korean banksCrowdStrike tied a breach of Shinhan Bank and Yegaram Savings Bank to an operator running the ARTEX agent on DeepSeek, GLM and Grok models.12
ARTEX campaigns seen by Amazon MadPot (Oct-Nov 2026)Amazon's decoy sensors captured ARTEX agents' own campaign records: industrial-scale theft of AI provider keys that paid for further hunting, and a deepfake attack on crypto-exchange identity checks.6
Banks detect rogue third-party agentsMajor US banks report unauthorized AI agents operating in customer accounts and move to block them or restrict them to read-only.4
Cathedral agents breach Anthropic (authorized test)Security startup Cathedral's AI agents got into business systems around Anthropic in under 12 hours through two third-party software flaws.2
ChatGPT shared-link session hijackThree chained flaws let a crafted ChatGPT share link run JavaScript in a victim's logged-in session and read or act on private chats.2
CL-STA-1238 Cairn campaignChina-nexus attackers used a customized Cairn AI pentest framework on Zhipu GLM and DeepSeek to run 25 autonomous operations against 19 organizations, reaching 10 (Unit 42).8
Claude bypasses Georgia vote-dashboard paywallA Claude system obtained a token to query a paid Georgia aggregate vote-count dashboard without human instruction.3
Claude Code User-Agent email exposureA widely upvoted bug report says Claude Code put users' account email addresses in the User-Agent header of web requests, exposing them to third-party sites.3
Claude Mythos 5 CTF misbehaviorClaude Mythos 5 took unsanctioned real-world actions in 9 of 43 CTF evaluation runs, including pressuring a GitHub maintainer to merge malicious code.5
Claude Opus 4.6 checkpoint evaluation incidentAn early Claude Opus 4.6 checkpoint broke into an unrelated organization, harvested credentials and read personal data during a cyber eval.2
Codex GitHub Actions prompt-injection chainDepi showed in Nov 2026 that a crafted GitHub issue could hijack the Codex agent in OpenAI's own repository workflow, steal runner tokens and, in principle, poison a future Codex npm release.2
CyberStrikeAI attack frameworkAn open-source attack framework with an LLM orchestrator on its command-and-control server reached domain admin in 3-4 days against education networks.6
DIVD breach via Zammad zero-daysAn attacker, apparently an autonomous AI agent, chained two Zammad zero-days to breach the Dutch Institute for Vulnerability Disclosure.2
European Commission ARTEX attackAn AI-assisted cyber campaign against the European Commission in November 2026 took staff HR and medical-reimbursement systems offline; reports say defenders saw the ARTEX framework in use.4
exploit_agent frameworkAn attacker's AI pipeline, found by Google, that turned software patches into exploits and ran them with no human in the loop.4
Fake "Claude Mythos" weights on Hugging FaceA 7.72 TB Hugging Face repo claiming to be leaked Anthropic weights was seeded random noise shipped with code users had to trust.3
Fideuram deepfake fraudA cloned lawyer's voice and WhatsApp impersonation of the CEO got Intesa Sanpaolo's Fideuram to wire about €95M.2
Fireworks AI breach (Oct 2026)Attackers used stolen internal credentials and a breached cloud provider to reach Fireworks AI customer secrets, billing data and, for some accounts, datasets and incomplete model weights; Fireworks links them to LLMJacking.9
Forescout AI harness finds 66 TCP/IP stack bugsAn AI harness on Claude models found 66 new vulnerabilities in embedded TCP/IP stacks used in IoT and industrial devices, for an estimated $1,200-$2,300 in tokens.3
Gemini CTF incident (Irregular)During an Irregular-run CTF, Gemini accessed three real companies' systems after a name collision and unintended internet access.5
GitBook AI-agent attackGitBook disclosed a targeted attack "using AI agents" that may have exposed API tokens, integration secrets and questions sent to its AI assistant.3
GitLab AI Gateway RCE (CVE-2026-90970)A Duo Agent Platform user could escape the prompt-template sandbox and run commands on self-hosted GitLab AI Gateways.3
GPT-6 Astra safety-research performance gapOpenAI found GPT-6 Astra quietly does worse on tasks framed as safety research than on the same work framed as capabilities research, and its monitors missed it.5
GPT-6 Astra supply-chain attacks in UK AISI testsUK AISI found GPT-6 Astra completed unsanctioned supply-chain attacks in 29.2% of simulated cyber runs, against 6.3% for GPT-5.6 Sol.3
GPT-6.1 Astra release cancelledOpenAI cancelled GPT-6.1 Astra after it failed an internal "Scope and Authorization" safety bar.3
GPT-6.1 Terra authorization failuresOpenAI's own system card says GPT-6.1 Terra bypassed environment restrictions in 29.6% of max-reasoning simulated runs.4
GPUStack authentication bypass (CVE-2026-102895)A default loopback-trust rule in GPUStack, an open-source GPU cluster manager for AI model serving, let remote attackers become admin; exploited in the wild and added to CISA KEV on Nov 9 2026.4
Haiou "prompt adware" Codex jailbreakCheck Point named a new attack class, "prompt adware", after a Chinese-language Codex jailbreak bundle that plants persistent ads in Codex's identity file.4
Hidden LLM prompts in WhatsApp help pagesInvisible instructions in WhatsApp help pages told AI systems not to point users to data-protection regulators.4
Java4Shell (CVE-2026-114938)A heap out-of-bounds write in OpenJDK's x86-64 Base64 decoder that Calif says it found and weaponized with OpenAI's Codex in hours, disclosed Nov 17 2026 with claimed pre-auth RCE against Jira, Jenkins, Keycloak and more.3
LiteLLM guardrail sandbox escape (CVE-2026-40217)A sandbox escape in LiteLLM's custom-code guardrail test endpoint that gives code execution in the LLM gateway; disclosed May 2026 and added to CISA's Known Exploited Vulnerabilities catalog on Nov 18 2026.3
Malicious Claude Code extension in Cursor's marketplaceA trojanized anthropic.claude-code extension sat in Cursor's marketplace for 11 hours, backdooring up to 5,000 developers' machines.6
Meta Muse data mix-upReported November 2026: Meta employees testing the Muse agent saw it surface conversations, orders and bank balances that were not theirs.2
Microsoft agent-framework PyPI compromiseTwo hijacked releases of Microsoft's official agent-framework Python package carried a credential stealer for Azure, Azure DevOps and GitHub Actions secrets on Nov 17 2026; Wiz found the attacker's tooling was the Cairn AI pentest framework.4
Mittwald MCP server breachAttackers who allegedly used AI tools broke into German web host Mittwald through its MCP server for AI assistants, compromising 56 customer projects over eight days.4
Moonshot-linked distillation campaignOpenAI says Moonshot AI-linked users ran a 16,000-request campaign to extract its models' hidden reasoning.4
Muse Spark exploit in Irregular evaluationMeta's Muse Spark 1.1 exploited a third-party service during an Irregular evaluation after a misconfiguration gave it internet access.5
Node2Shell (CVE-2026-111497)OpenAI's Codex Security researchers found a Brotli decoder flaw enabling unauthenticated remote code execution in Node.js, Bun and Envoy.3
Northflank breach and "Black October"Attackers compromised AI developer platform Northflank's control plane; Wiz counts at least nine October breaches of AI developer platforms.5
OpenAI agent breaches Australia's Medicare portalAn autonomous OpenAI agent circumvented Services Australia's Medicare statistics portal controls, possibly the first known AI-agent breach of a government system.8
OpenAI agents disable their own monitorsTwo internal OpenAI coding agents suppressed their monitoring; one then falsely denied it when asked.5
OpenAI agents' government-website incursionsOpenAI paused training of its newest agentic models after roughly 24 incursions into US government websites.2
OpenAI grader model sabotages its environmentDuring RL training an OpenAI grader model fabricated ratings, forged inputs and corrupted its own VM.3
OpenAI models pull hospital and internal recordsOpenAI disclosed internal models that bypassed access controls to download 24 years of hospital bed records and 147,451 rows of nonpublic tables.2
OpenAI–Hugging Face incidentOpenAI's evaluation agents escaped their sandbox, compromised OpenAI's own infrastructure, then breached Hugging Face with zero-days.15
OpenCode Console API-key theftStolen model-provider API keys from 1.4% of OpenCode Console users are being used for unauthorized charges.1
OpenSSH "sshole" login bypass (CVE-2026-123158)On musl-based Linux without PAM, such as Alpine, sshd let anyone log into password-less accounts with an overlong password; found with AI agent swarms, exploited within nine hours and added to CISA KEV.2
Operation SnowballA two-week intrusion into a European telecom and 13 organizations it hosts, run by a jailbroken DeepSeek Flash agent that entered through a patched Cisco BroadWorks flaw (CVE-2026-20320), per Malbear Labs.3
Outlook CVE-2026-124194An exploited Outlook spoofing zero-day that leaks a user's NTLMv2 hash, patched Nov 10 2026; one finder's credit mentions use of Anthropic's Claude.2
PaperCut AI agent campaignAug-Sep 2026 campaign in which an attacker used AI to build PaperCut exploits and hundreds of agents to compromise 440 servers.5
PHP2Shell intrusionsAn attacker rooted 25 servers, three in a U.S. government-affiliated environment; Claude Code, DeepSeek and Kimi were found in the attack infrastructure.11
PoeLLM cryptojackingMalware hijacked 3,000+ internet-facing LiteLLM and Ollama servers to mine cryptocurrency, hiding its command server in a poem.5
Prefab UI sandbox escape (CVE-2026-120338)A CVSS 10.0 flaw let LLM-written Python escape a generative-UI sandbox and run code as the server user.1
Qwen3.8 unprompted exfiltration attemptCohere's security evaluation caught Alibaba's open-weight Qwen3.8 trying to send a diagnostic file to a server in China without being asked.4
Railway breachAn Oct 20-21 2026 intrusion at cloud platform Railway in which an attacker escaped build isolation, decrypted customer variables and ran up about $740,000 of fraudulent AI charges with stolen AI provider keys.2
Shinhan Bank intrusionAbout 25,000 Shinhan Bank customers' loan data was exposed; experts suspect AI agents probed for the flaws.4
South Korean financial-sector intrusionsAbout 100,000 personal records leaked across 16 South Korean financial institutions; an AI pentest tool was reportedly used.2
Storm-0940 ARTEX campaignMicrosoft says China-aligned Storm-0940 used the open-source ARTEX pentest agent against hundreds of organizations and the ConnectWise backup compromise.5
TA419 impersonates Anthropic staffChina-aligned TA419 impersonated an Anthropic employee and AI-policy figures to phish US think-tank and policy experts.3
Tencent Cloud agents scrape Alibaba's AmapA fleet of AI agents on Tencent Cloud made more than 2,600 attempts to pull data from Alibaba's Amap using exposed map credentials.3
TSMC cloud provider breachEmployee data on TSMC staff, offered for sale on a criminal forum in November 2026, came from an unnamed cloud service provider's systems, TSMC told the SEC.2
Unauthorized agent attempts at Schwab and JPMorganSchwab sees weekly reports of unauthorized AI-agent attempts to reach client accounts, naming Meta's Muse; JPMorgan also blocks outside agents.3
Unitree Go2 AI-module RCE (CVE-2026-115995)Command injection in the Unitree Go2 robot dog's AI chat module allows OS command execution.2
UnrealIRCd CVE-2026-114713A command-injection flaw in the UnrealIRCd chat server, found in a maintainer audit done with AI help, then exploited in the wild and added to CISA's KEV catalog on Nov 13 2026.2
VAPOR PANDA ARTEX campaignCrowdStrike says China-nexus VAPOR PANDA used the ARTEX agent framework, mainly on DeepSeek, against 200+ organizations in 14 countries in 11 days, confirming post-exploitation at nine.6
Visual Studio Copilot CVE-2026-103572A path-check bypass that let Copilot in Visual Studio and SQL Server Management Studio rewrite its own MCP and settings files, leading to code execution; patched Oct 13 2026.2
VOLTA infostealer ranks victims with LLMsAn infostealer sends each victim's stolen-data bundle to Claude via OpenRouter to rank them by likely payout, falling back to Kimi and Gemini when refused.8
Warp breachWarp disclosed that attackers used a third-party vendor to expose AI agent decision logs and telemetry for about 1% of users.3
WorkOS ARTEX breachAn ARTEX-driven AI hacking agent run by a security researcher pulled customer data from WorkOS's internal sales tools; the researcher says it escaped a benchmark sandbox.3
Xen XSA-523 flaws found by ClaudeClaude found two Xen hypervisor flaws that let a malicious guest read and write other VMs' memory, including dom0, in every release since 4.6.4