Tim McAllister · AI incident watch
See where the reports lead.
Explore documented affected regions, then open an incident to see what happened and what to learn from it.
Register reviewed . Location review .
The global picture
AI incidents, in context.
Explore affected regions. Select a marker to read the reports behind it.
Drag to pan. + / − to zoom, arrow keys to pan. Ctrl/⌘ + scroll to zoom. Use the region list on touchscreens.
Regions are approximate, not exact incident locations. Markers show documented affected areas, not an attacker’s location. Nearby regions cluster as you zoom out; counts show distinct reports, and a report may appear in more than one region.
Read the record
All affected regions
62 reports · newest first
- Observed incidentAoki infostealer adds AI assistant credentials to its targets
Location not established
- Observed incidentCERT-In reports unauthorized agent access to Indian organizations
India
- Observed incidentCL-UNK-1211 uses suspected AI-generated scripts in government intrusions
Location not established
- Vulnerability / researchCounterfeit-policy research steers Jev security judgments
Location not established
- Observed incidentOpenAI evaluation model changes live systems while researching a dashboard
Location not established
- Observed incidentOpenAI training model routes blocked requests through an external browser
Location not established
- Observed incidentOpenAI training model retrieves government-site secrets and bypasses authorization
Location not established
- Observed incidentOpenAI research model spends an unauthorized test-network wallet's funds
Location not established
- Observed incidentAIR reports live credential exposure from incorrect Render skill guidance
Location not established
- Vulnerability / researchSneakPeek exposes local secrets through Claude's preview pane
Location not established
- Vulnerability / researchXBOW demonstrates an AI-assisted Chrome exploit chain
Location not established
- Observed incidentAttackers target Langflow MCP configuration for code execution
Location not established
- Observed incidentClaude evaluations and internal work exceed real-site permissions
United States
- Vulnerability / researchAgentCorruption demonstrates broad AgentCore permissions risk
Location not established
- Observed incidentPoeLLM botnet compromises exposed AI services
United States · Western Europe
- Vulnerability / researchGhostSquatting exposes abandoned names in AI instructions
Location not established
- Observed incidentWikimedia finds unauthorized agent edits and failed proxy attempts
Location not established
- Observed incidentARTEX-linked campaign targets South Korean financial organizations
South Korea
- Observed incidentCanada responds to reports of suspected AI-agent probing
Canada
- Observed incidentOpenAI research model publishes a GitHub token
Location not established
- Observed incidentData-retrieval agents attempt intrusions at public data sites
United States · Australia
- Observed incidentGTG-10007 uses agent swarms for intrusions
China · Southeast Asia
- Observed incidentGTG-20006 automates espionage with Claude
Ukraine · Europe · Middle East · Southeast Asia · North Africa
- Observed incidentAnthropic discloses another evaluation intrusion from January
Location not established
- Observed incidentAI-orchestrated PaperCut campaign compromises hundreds of servers
United States · United Kingdom · France · Spain · Canada · Belgium · Portugal · Australia · Germany · Switzerland · Italy · Taiwan · Singapore · Netherlands · South Africa · Sweden · Brazil · Malaysia · Denmark · Ireland · New Zealand · Argentina · India · Cambodia · Chile · Finland · Greece · Japan · Puerto Rico · Austria · Botswana · Bulgaria · China · Colombia · Ecuador · Estonia · Kazakhstan · Lithuania · Mexico · Namibia · Nigeria · Pakistan · Philippines · Poland · Romania · Saudi Arabia · Sri Lanka · Zimbabwe
- Observed incidentAI-assisted attacker builds a credential-harvesting operation in hours
Location not established
- Observed incidentUK AISI reports unauthorized live actions during cyber tests
Location not established
- Observed incidentMisconfigured Claude evaluations reach three real organizations
Location not established
- Vulnerability / researchAgentForger turns a crafted link into a deployed ChatGPT agent
Location not established
- Observed incidentOpenAI evaluation models reach Hugging Face production systems
Location not established
- Vulnerability / researchUntrusted resume input enables command execution through Claude Python SDK
Location not established
- Vulnerability / researchGhostApproval hides coding agents' true file-write destinations
Location not established
- Observed incidentJADEPUFFER chains AI-assisted intrusion and destructive extortion
Location not established
- Observed incidentAttackers probe LiteLLM guardrail testing for code execution
Location not established
- Vulnerability / researchGemini workflow research connects prompt injection to excessive cloud privileges
Location not established
- Observed incidentGoogle disrupts planned exploitation of an AI-assisted 2FA bypass
Location not established
- Vulnerability / researchCurseChain carries malicious instructions into a later Cursor task
Location not established
- Observed incidentClaude-assisted crypto project adds a PromptMink dependency
Location not established
- Vulnerability / researchAntigravity native tool bypass permits code execution
Location not established
- Vulnerability / researchComment and Control exposes coding-agent credentials
Location not established
- Vulnerability / researchPipeLeak redirects Agentforce custom actions toward data exfiltration
Location not established
- Observed incidentMalicious LiteLLM PyPI releases carried credential-stealing code
Location not established
- Observed incidentEvilTokens combined account compromise with AI-assisted fraud preparation
United States · Canada · United Kingdom · Australia · India · France
- Vulnerability / researchAI-led testing uncovered a vulnerability in McKinsey's Lilli
Location not established
- Vulnerability / researchClawJacked let a website take over a local OpenClaw agent
Location not established
- Observed incidentAI-assisted actor compromised hundreds of FortiGate devices
South Asia · Latin America · Caribbean · West Africa · Northern Europe · Southeast Asia
- Observed incidentCompromised Cline publishing token caused unauthorized OpenClaw installation
Location not established
- Vulnerability / researchRoguePilot demonstrated repository takeover through Copilot in Codespaces
Location not established
- Observed incidentSnyk confirmed malicious packages in the agent skills ecosystem
Location not established
- Observed incidentMoltbook database misconfiguration exposed agent tokens and private data
Location not established
- Observed incidentBizarre Bazaar monetized exposed AI endpoints
Location not established
- Vulnerability / researchShareLeak let form submissions redirect Copilot Studio agents
Location not established
- Vulnerability / researchReprompt turned a Copilot link into a data-exfiltration chain
Location not established
- Vulnerability / researchBodySnatcher exposed ServiceNow AI workflows through an authentication flaw
Location not established
- Vulnerability / researchDocker repository metadata redirected Ask Gordon into data exfiltration
Location not established
- Vulnerability / researchIDEsaster chained AI file writes into IDE execution and data leakage
Location not established
- Observed incidentShai-Hulud 2.0 malicious releases reached the mcp-use package
Location not established
- Observed incidentMixpanel breach exposed limited OpenAI user analytics
Location not established
- Vulnerability / researchAntigravity browser subagent leaked credentials in a researcher demonstration
Location not established
- Observed incidentAnthropic disrupts a Claude-assisted espionage campaign
Global / cross-border
- Observed incidentGoogle documents Gemini misuse for intrusion support and experimental malware
Location not established
- Vulnerability / researchHidden webpage instructions exposed Opera Neon account data
Location not established
What the map shows
Markers represent approximate country or region centers, based on the affected or targeted locations named in the sources. They do not identify exact incident sites. Attempts remain attempts, and research remains research in the linked reports. Attacker origin, a vendor’s headquarters, and a researcher’s location are not treated as affected locations. Reports can span more than one region; regional counts overlap.
Some sources do not disclose where an incident occurred. Those reports remain available in the unlocated list. Broad international campaigns without a documented regional breakdown are listed separately. This curated register is not a measure of incident prevalence by country.
Location evidence
- CERT-In reports unauthorized agent access to Indian organizations
India
- CERT-In advisory CIAD-2026-0049 — The stored title and parties identify Indian organizations as the targets. The victims remain unnamed; this is country-level coverage only.
- Claude evaluations and internal work exceed real-site permissions
United States
- Anthropic investigation of unintended model actions — The report identifies affected federal, state and local US government websites, including a Philadelphia Police Department tip form. Other unnamed cases are not located by this marker.
- PoeLLM botnet compromises exposed AI services
United States, Western Europe
- Black Lotus Labs: Canto Incognito — Lumen places the main concentrations of compromised victim servers in the United States and Western Europe. These are affected-host regions, not the operator location; the full victim population is broader.
- ARTEX-linked campaign targets South Korean financial organizations
South Korea
- CrowdStrike ARTEX investigation — The primary investigation identifies financial organizations in South Korea as campaign targets. Country-level placement does not identify any particular organization or site.
- Canada responds to reports of suspected AI-agent probing
Canada
- Canadian government statement — The statement identifies suspicious requests targeting Government of Canada public websites. It reports no indication of compromised government systems; the marker denotes targeted services, not a proven breach.
- Data-retrieval agents attempt intrusions at public data sites
United States, Australia
- Transluce public-data investigation — The three intrusion-attempt targets are Data USA, the University of New Mexico digital library and the Australian Institute of Health and Welfare. US and Australian service locations are shown; the reported probes did not establish nonpublic-data access.
- GTG-20006 automates espionage with Claude
Ukraine, Europe, Middle East, Southeast Asia, North Africa
- Anthropic September 2026 report: GTG-20006 — This section describes Ukrainian targets, wider European and Middle Eastern targeting, a Southeast Asian maritime-government target, and an intrusion into a North African government authority. Only Ukraine is named as a victim country.
- GTG-10007 uses agent swarms for intrusions
China, Southeast Asia
- Anthropic September 2026 report: GTG-10007 — The GTG-10007 section places hands-on intrusions primarily against domestic Chinese victims and describes citizen-record retrieval from a Southeast Asian government agency. Operator cities and separate reconnaissance regions are not plotted.
- AI-orchestrated PaperCut campaign compromises hundreds of servers
United States, United Kingdom, France, Spain, Canada, Belgium, Portugal, Australia, Germany, Switzerland, Italy, Taiwan, Singapore, Netherlands, South Africa, Sweden, Brazil, Malaysia, Denmark, Ireland, New Zealand, Argentina, India, Cambodia, Chile, Finland, Greece, Japan, Puerto Rico, Austria, Botswana, Bulgaria, China, Colombia, Ecuador, Estonia, Kazakhstan, Lithuania, Mexico, Namibia, Nigeria, Pakistan, Philippines, Poland, Romania, Saudi Arabia, Sri Lanka, Zimbabwe
- GreyNoise PaperCut victim-country table — The victim-country table lists affected instances in these 48 countries and territories. The separate attacker exclusion list is not used. Each marker represents one campaign report associated with that area, not a victim count.
- EvilTokens combined account compromise with AI-assisted fraud preparation
United States, Canada, United Kingdom, Australia, India, France
- Microsoft EvilTokens disruption report — Microsoft identifies the United States, Canada, United Kingdom, Australia, India and France as the highest concentrations of victim activity within a worldwide campaign. These six countries are documented concentrations, not complete coverage.
- AI-assisted actor compromised hundreds of FortiGate devices
South Asia, Latin America, Caribbean, West Africa, Northern Europe, Southeast Asia
- Amazon FortiGate geographic distribution — Amazon reports compromised-device concentrations in South Asia, Latin America, the Caribbean, West Africa, Northern Europe and Southeast Asia, among other regions. More than 55 countries were affected; no country list is provided.
- Anthropic disrupts a Claude-assisted espionage campaign
Broad international scope; individual locations not assigned.
- Anthropic AI espionage disclosure — The report describes roughly 30 global organizational targets without a usable victim-country breakdown. No country is assigned from the suspected actor attribution.
Map outlines: Natural Earth, public-domain 1:110m data. Register scope and methodology.