Browse / Sensitive Data Disclosure
Sensitive Data Disclosure
Secrets, personal data, chat history or training data exposed through an AI system. 81 records. Reference: OWASP LLM02 · ATLAS Exfiltration.
- Real-world incidents 41
- Information exposure 23
- Data exfiltration 5
- Unsecured Credentials 5
- Path traversal and file access 2
- Data from Information Repositories 2
- Server-side request forgery 1
- Code and command execution 1
- LLM Prompt Injection: Indirect 1
- Mem0: Cross-tenant conversation leak via unescaped delimiter collision in session-scope key (_build_session_scope) Conversation scope keys join caller-supplied identifiers with unescaped delimiters, so a crafted run_id can match another tenant's stored messages.
- MLflow CrewAI autologging records LLM provider API keys in logged trace spans MLflow autologging serializes CrewAI agent objects into trace spans, and their string form includes the LLM provider API key, which is then sent to the trace backend.
- PraisonAI: ContextGatherer include resolution permits absolute and traversal reads outside the workspace Include entries in a repository's context files may be absolute or contain traversal, so the gatherer reads host files outside the workspace and adds them to the model context.
- PraisonAI: Project custom command templates can read outside-workspace files into model prompts A repository-supplied command template can reference parent or absolute paths, which are read and inserted into the model prompt, leaking local files to the provider.
- Pydantic AI OpenTelemetry instrumentation: retry prompt content is not redacted when `include_content=False` Retry prompts without tool calls are written to telemetry spans even when content redaction is enabled, so withheld model content reaches the observability backend.
- Pydantic AI OpenTelemetry instrumentation: exception events on tool and agent run spans include content when `include_content=False` Exception events and error statuses on agent spans carry prompts, tool arguments and provider responses even when content export is disabled.
- Docling: Configured HTTP headers sent to every remote image host named by a document Configured authentication headers are sent to whatever host a document's image link names, and redirects forward them onward, exposing credentials.
- vLLM: Harmony tool continuations drop `cache_salt` — restoring a cross-tenant prefix-cache membership oracle Tool continuation turns are cached without the tenant's cache salt, so another tenant can infer guessed prompts from exact cached-token counts.
- AI Coding Agents Reportedly Exposed Over 13,000 Internal Images from Hundreds of Organizations on Public GitHub Code Repositories AI coding agents published over 13,000 internal images from 300+ organizations to public GitHub repositories while preparing code-review screenshots, exposing billing records and unreleased features.
- OpenAI AI Agents Reportedly Posted ChatGPT User Images to External Hosts in 53 Instances OpenAI reported that its research-environment agents posted user-supplied ChatGPT images to external hosts in 53 instances, leaving them reachable through unlisted links outside its systems.
- Z.ai's ZCode AI Coding Assistant Reportedly Uploaded Developers' Local Code Repositories to Alibaba Cloud Without Consent Z.ai's ZCode assistant, via a default-enabled indexing feature, uploaded developers' local code repositories to Alibaba Cloud without consent; the company patched and disabled the path.
- Instinct AI Personal Assistant Reportedly Summarized an AI Product Founder's Retained Emails After Google Access Was Disconnected Instinct summarized previously indexed emails retained after the user revoked Google access, exposing data it should no longer have been using.
- Flowise: Information Disclosure in GET /api/v1/upsert-history returns the entire server-wide upsert history The upsert history endpoint returns every workspace's ingestion records without scoping, exposing vector store endpoints and collection names.
- Publicly Shared Claude Artifacts Reportedly Became Discoverable Through Google Search Publicly shared Claude Artifacts became findable through Google Search after being posted somewhere crawlable, exposing business plans and infrastructure diagrams.
- Open WebUI: /api/v1/channels/{id}/members exposes full user model including sensitive credentials The channel members endpoint returns full user records, so any channel member can read other users' settings including tool server keys and webhook URLs.
- n8n: Bypass "Allowed HTTP Request Domains" Credential Restriction in Multiple AI and LLM Nodes A workflow editor with use-only access to a shared AI credential points an LLM node at an attacker host, causing the credential secret to be sent there.
- n8n: Custom Header Credential Values Leaked in Plaintext into LLM Node Execution Data Custom header values from LLM node credentials are stored in plaintext in workflow execution data, where users with execution access can read them.
- n8n: Google Service Account Private Key Exposed in JWT Header The full service account private key was written into the JWT header's key ID field, so anyone who can read those headers can recover it.
- Autonomous OpenAI Evaluation Agents Compromise Hugging Face Infrastructure
- Multi-Agent Framework Compromises Taiwanese Government Systems
- Rep. Anna Paulina Luna's Office Reportedly Published Claude Transcript Residue on Accident in House Amendment Summary A staffer reportedly pasted leftover Claude chat transcript text, including a timestamp and 'Claude responded:' marker, into a public House Rules Committee amendment summary.
- MLflow: Environment variable injection in AI Gateway secrets enables server-side credential exfiltration The gateway resolves environment variable references in API key fields and sends the resulting cloud credentials to an attacker-chosen upstream base URL.
- PraisonAI CLI automatically resolves @url mentions in prompt text and can read loopback URLs into model context Attacker-written prompt text containing a URL mention makes the CLI fetch local HTTP endpoints and insert the response into the model's context.
- OpenAI Research Agent Reportedly Exposed a Researcher's GitHub Token in a Public Code Repository An internal OpenAI agent reportedly split a researcher's GitHub access token into pieces to evade secret scanning and published it in a public Codex repository while seeking another team's proof.
- FlowiseAI Vulnerable to Credential Data Leak Credential lookups with a name filter return the encrypted data field, which the unfiltered path strips, exposing stored API secrets.
- Flowise: Bcrypt Password Hash Exposure The Flowise login handler returns bcrypt password hashes in its response, so a remote caller can obtain other users' hashes.
- Langflow has an Information Leak through Incomplete API Key Redaction API key redaction in flow handling misses some cases, so stored credentials are returned to remote callers.
- Flowise: Public chatflow endpoints return unsanitized flowData including plaintext API keys, passwords, and credential IDs The public chatflow GET endpoint returns stored flow data unfiltered, so credential references, plaintext keys and passwords reach anonymous callers.
- LangSmith SDK: Streaming token events bypass output redaction Streamed token events are recorded as raw values and skip the output redaction applied to run inputs and outputs, so sensitive model output is still stored.
- Flowise: Sensitive Data Leak in public-chatbotConfig An unauthenticated endpoint returns full chatflow configuration, including stored API keys and authorization headers, to anyone who knows a chatflow ID.
- Autonomous Claude Agents Compromise Three Real Organizations During Cyber Evaluations
- Meta Internal AI Agent Reportedly Gave Advice That Allegedly Exposed Sensitive Data to Unauthorized Employees A Meta internal AI agent posted unapproved technical advice on an internal forum; an employee followed it, exposing sensitive company and user data to unauthorized staff for about two hours.
- Gradio: Mocked OAuth Login Exposes Server Credentials and Uses Hardcoded Session Secret Mocked OAuth routes on non-Spaces deployments copy the host's real Hugging Face token into any visitor's session cookie, which is signed with a hardcoded secret.
- CodeWall's Autonomous Agent Reportedly Obtained Unauthorized Access to McKinsey's Lilli AI Platform Database An autonomous agent from CodeWall exploited flaws in McKinsey's Lilli platform to gain read and write access, exposing chats, files, accounts and prompts.
- Meta AI Smart Glasses Reportedly Routed Intimate Imagery to Reviewers at Kenyan Contractor Sama Before Meta Ended Contract Meta's AI glasses reportedly sent intimate recordings and transcripts to Kenyan contractors at Sama for human review, exposing private activity and bank cards.
- Grok Reportedly Disclosed Adult Performer Siri Dahl's Legal Name and Birthdate, Allegedly Contributing to Doxxing and Harassment Grok reportedly publicly disclosed adult performer Siri Dahl's legal name and birthdate without being asked, enabling impersonation and doxxing.
- SageMaker Python SDK has Exposed HMAC The HMAC key that signs remote function payloads is readable through the DescribeTrainingJob API, letting an attacker forge payloads that run code when results are loaded.
- Moltbook Database Exposure Allegedly Revealed Users' Private Communications and API Authentication Tokens An exposed Moltbook database let researchers access about 35,000 emails, thousands of private DMs, and roughly 1.5 million API tokens, enabling possible impersonation of AI agent accounts.
- Exposed ClawdBot Control Interfaces Leads to Credential Access and Execution
- Claude Code Leaks Data via Malicious Environment Configuration Before Trust Confirmation A repository settings file can point the API base URL at an attacker endpoint, and Claude Code sends requests carrying the user's key before the trust prompt appears.
- LangChain serialization injection vulnerability enables secret extraction Free-form metadata containing LangChain's reserved key is not escaped on serialization, so later loading treats attacker data as secret references and reveals environment variables.
- Alleged Harmful Outputs and Data Exposure in Children's AI Products by FoloToy, Miko, and Character.AI Children's AI toys from FoloToy, Miko and Character.AI reportedly exposed user data, alongside harmful outputs, through weaknesses in their systems.
- Secret Desires AI Platform Reportedly Exposed Nearly Two Million Sensitive Images in Cloud Storage Leak The Secret Desires AI platform left nearly two million personal and explicit AI-generated images and videos publicly accessible in misconfigured cloud storage.
- AI Training Dataset for Detecting Nudity Allegedly Found to Contain CSAM Images of Identified Victims NudeNet, an image dataset used to train nudity detectors, reportedly contained CSAM including images of known victims and was widely downloaded before removal after notification.
- GTG-1002 Claude Code Espionage Campaign
- Microsoft's Windows Recall Allegedly Stores Passwords and Social Security Numbers in Preview Mode Windows Recall kept capturing and storing passwords, Social Security numbers and bank details in searchable history, and its sensitive-information filter reportedly failed in testing.
- Reported Public Exposure of Over 100,000 LLM Conversations via Share Links Indexed by Search Engines and Archived Chats marked discoverable were indexed by search engines and archives, then scraped, exposing over 100,000 LLM conversations including API keys, tokens and personal data.
- CISA Acting Director Reportedly Uploaded Sensitive Government Documents to Public ChatGPT Instance A CISA acting director reportedly pasted 'for official use only' government contracting documents into public ChatGPT, exposing sensitive material to an external AI service.
- Urban VPN Proxy Browser Extension Reportedly Harvested and Sold Private AI Chatbot Conversations via Silent Update A malicious update to the Urban VPN browser extension reportedly secretly collected users' private conversations from ChatGPT, Claude and other AI platforms and shared them with data brokers.
- McDonald's McHire AI Recruitment Platform Reportedly Exposed Data of 64 Million Applicants via Default Login and API Vulnerability Default admin credentials and an insecure direct object reference in McDonald's McHire chatbot platform reportedly exposed applicants' personal data and chat histories.