AI leaks happen every single day.

API Key LeaktechnologyLiteLLM

Today

Delve did the security compliance on LiteLLM, an AI project hit by malware

LiteLLM, an open-source AI project used by millions, was compromised by credential harvesting malware that likely exfiltrated API keys and authentication tokens. This represents a significant supply-chain compromise affecting widespread AI infrastructure.

“LiteLLM offers an AI open source project used by millions that was infected by credential harvesting malware”

Preventable by Bleep

A DLP proxy could detect suspicious API key extraction attempts or unusual API usage patterns from compromised systems, but cannot prevent malware execution that occurs at the source code/binary level.

Credential ExposureChatGPTClaude

Yesterday

Paid AI Accounts Are Now a Hot Underground Commodity

Paid AI service accounts (ChatGPT, Claude, etc.) are being stolen and resold on underground markets as part of the cybercrime supply chain. This represents unauthorized access to premium AI accounts and potential misuse of services for malicious purposes.

Preventable by Bleep

A DLP proxy could detect unauthorized API key usage and suspicious patterns in AI service access, but cannot prevent credential theft that occurs outside the organization or on the dark web.

Credential ExposuretechnologyLiteLLM

Yesterday

Checkmarx KICS Code Scanner Targeted in Widening Supply Chain Hit

Yesterday

How AI Coding Tools Crushed the Endpoint Security Fortress

Securitytechnology

Credential ExposuretechnologyLiteLLM

TeamPCP Backdoors LiteLLM Versions 1.82.7–1.82.8 via Trivy CI/CD Compromise

TeamPCP compromised LiteLLM, a popular Python package for LLM interactions, injecting malicious code including a credential harvester, Kubernetes lateral movement toolkit, and persistent backdoor in versions 1.82.7 and 1.82.8. This directly impacts AI tool users and could lead to widespread API key exposure and infrastructure compromise.

Preventable by Bleep

A DLP proxy could potentially detect if compromised credentials extracted by the backdoor are being exfiltrated to external services, but the compromise happens at the library level before proxy inspection.

Credential Exposuretechnology

GitHub 'OpenClaw Deployer' Repo Delivers Trojan Instead

An AI-assisted campaign distributed over 300 poisoned packages through GitHub via the 'OpenClaw Deployer' repository, targeting developer tools and gaming assets. This represents a large-scale supply chain attack leveraging AI to enhance distribution capabilities.

“An AI-assisted campaign is spreading more than 300 poisoned packages for diverse assets ranging from developer tools to game cheats”

Preventable by Bleep

A DLP proxy could potentially detect if these trojanized packages attempt to exfiltrate stolen credentials or sensitive data to external services, but the initial infection happens at installation.

Corporate DatatechnologyClaude CodeCursor

AI coding tools have broad filesystem and network access

AI coding tools like Claude Code, Cursor, and Copilot have broad filesystem and network access without user visibility into which files are being read or what data is being transmitted. The creator developed Agent Shield to monitor and control what these tools access on local machines, highlighting a significant transparency and data leakage risk.

Preventable by Bleep

A DLP proxy could monitor and block network transmissions of sensitive files (like .env, .ssh) being sent to external AI services, but cannot directly control the OS-level filesystem access that these local tools have. Client-side monitoring would be more effective than network-level DLP.

OpenAI rolls out ChatGPT Library to store your personal files

PII ExposureChatGPT

Corporate DatatechnologyMeta AI

6d ago

Meta AI agent's instruction causes large sensitive data leak

6d ago

AI Conundrum: Why MCP Security Can't Be Patched Away

A researcher at RSAC 2026 highlights that MCP (Model Context Protocol) introduces architectural security risks into LLM environments that cannot be patched away. This addresses systemic vulnerabilities in how LLMs integrate with external tools.

Preventable by Bleep

A DLP proxy could monitor and restrict data flows between LLMs and external tools accessed via MCP, potentially mitigating some risks, but architectural vulnerabilities in the protocol itself cannot be prevented by client-side proxies.

Securitytechnology

I built a Chrome extension that warns when you paste secrets into AI tools

A Chrome extension tool warns users when they attempt to paste secrets (API keys, credentials) into AI chat tools. This addresses a known security risk where users accidentally share sensitive credentials with AI services.

Preventable by Bleep

A DLP proxy monitoring AI traffic could identify and block credential and secret pasting into AI tools, providing similar functionality to this extension but at a network level with broader coverage.

Credential Exposure

Credential ExposuretechnologyClaude Code

How Ceros Gives Security Teams Visibility and Control in Claude Code

MCP servers mass-forked and republished – supply-chain attack vector

An organization systematically forked hundreds of MCP (Model Context Protocol) servers without author consent and republished them under different npm/PyPI scopes through their own marketplace. This represents a supply-chain attack vector that could allow injection of malicious code into AI tool dependencies used by developers.

IP ThefttechnologyClaude

'Claudy Day’ Trio of Flaws Exposes Claude Users to Data Theft

'Claudy Day' describes a trio of flaws in Claude that enable prompt injection and data theft attacks through Google search integration, creating a full attack chain that could compromise enterprise networks and user data. This is a direct AI service vulnerability enabling unauthorized data exfiltration.

Preventable by Bleep

A DLP proxy could potentially detect exfiltrated data being returned from Claude or sent to attacker-controlled endpoints, but cannot prevent prompt injection attacks occurring within the AI model itself or detect data being stolen within the model's response context.

Prompt InjectionClaude

Shadow AI is everywhere. Here’s how to find and secure it.

Corporate Data

Credential Exposuretechnology

Show HN: OpenLegion – AI agent fleet with container isolation and vault proxy

Credential ExposuretechnologyOpenClawAutoGen

We saw how 30 AI agent projects handle authorization-93% use unscoped API keys

Security audit of 30 popular AI agent projects reveals systemic authorization failures: 93% use unscoped API keys, 0% implement per-agent cryptographic identity, and 100% lack per-agent revocation capabilities. This represents widespread credential exposure risk across major AI frameworks.

Preventable by Bleep

A DLP proxy could detect and block unscoped API keys in agent configurations, alert on credential passing in API calls, and enforce scoped token injection at the proxy layer.

Prompt InjectionOpenClawClawdbot

OpenClaw AI Agent Flaws Could Enable Prompt Injection and Data Exfiltration

China's CNCERT issued a formal security warning about OpenClaw, an open-source AI agent platform, citing inherently weak default security configurations that enable prompt injection attacks and data exfiltration. This represents a documented real-world vulnerability affecting autonomous AI agents used in self-hosted deployments.

Preventable by Bleep

A DLP proxy could monitor and block data exfiltration attempts from the compromised AI agent, but cannot prevent the underlying prompt injection vulnerability or weak security misconfigurations within the platform itself.

Analyzing OpenClaw's 3-layer defense against prompt injection

Analysis reveals that OpenClaw and similar AI execution engines are vulnerable to prompt injection attacks that can lead to data exfiltration and full agent hijacking when agents have access to shells, databases, and web tools. Current defense mechanisms using regex blacklists and XML tagging are ineffective against semantic variations of injection attacks.

Preventable by Bleep

A DLP proxy could detect and block exfiltrated data being sent externally, but cannot prevent prompt injection vulnerabilities or agent hijacking that occurs within the AI execution environment itself.

Prompt InjectionOpenClaw

Accidental SharingtechnologyOpenClaw

Show HN: Privacy Mask – prevent secrets leaking to AI agents

AI-generated Slopoly malware used in Interlock ransomware attack

Security

Hive0163 Uses AI-Assisted Slopoly Malware for Persistent Access in Ransomware Attacks

The Hive0163 threat actor is using AI-generated malware called Slopoly for persistent access in ransomware attacks. This demonstrates how threat actors are weaponizing AI tools to develop malware frameworks faster, representing a significant AI security concern beyond traditional data leakage.

Security

Prompt InjectionPerplexity Comet

Researchers Trick Perplexity's Comet AI Browser Into Phishing Scam in Under Four Minutes

Researchers demonstrated that Perplexity's Comet AI browser can be tricked into phishing and scam traps by manipulating the AI's reasoning to lower security guardrails. The attack takes under 4 minutes, showing vulnerability of agentic AI browsers to adversarial prompts.

Preventable by Bleep

A DLP proxy could detect suspicious URLs or phishing domains being accessed by the Comet browser, but cannot prevent the prompt injection attack itself which occurs at the model level.

What Boards Must Demand in the Age of AI-Automated Exploitation

Compliance

API Key LeaktechnologyGPT-4

GPT-4 leaks its own API internals through training data exposure

GPT-4 was found to leak OpenAI's EPHEMERAL_KEY credential from its Realtime API through training data exposure when subjected to security tests, with a 75% leak rate across different prompts. This represents a definitive incident where sensitive API credentials were exposed through a model's training data.

How to Stop AI Data Leaks: A Webinar Guide to Auditing Modern Agentic Workflows

A webinar guide addressing the critical data leakage risks of AI agents that can autonomously send emails, move data, and manage software, creating a 'back door' for hackers. The 'invisible employee' concept highlights how AI agents with unsupervised access represent a massive data exfiltration vector.

“While these agents make work faster, they also open a new 'back door' for hackers”

Preventable by Bleep

A DLP proxy could monitor and block sensitive data flows from AI agents to external services, though it cannot prevent local file access or actions by client-side agents.

PII Exposure

New "LeakyLooker" Flaws in Google Looker Studio Could Enable Cross-Tenant SQL Queries

Corporate Data

Prompt Injectiontechnology

Energy-based Model (EBM) for enterprise AI security Ship it or keep tuning?

Credential ExposuretechnologyClaude

'InstallFix' Attacks Spread Fake Claude Code Sites

'InstallFix' is a cyberattack campaign using malvertising and ClickFix techniques to distribute fake Claude code sites and steal credentials from developers. The attack exploits the popularity of AI coding assistants by creating counterfeit installation pages.

Preventable by Bleep

A DLP proxy could detect when developers attempt to send credentials or sensitive information to fraudulent Claude sites by monitoring the destination URLs and blocking suspicious uploads.

Are We Ready for Auto Remediation With Agentic AI?

An article discussing organizational readiness for automated risk remediation using agentic AI. This is a governance and preparedness article rather than a report of an actual data leakage incident.

Security

Credential Exposuretechnology

Malicious npm Package Posing as OpenClaw Installer Deploys RAT, Steals macOS Credentials

A malicious npm package '@openclaw-ai/openclawai' masquerading as an OpenClaw installer deploys a remote access trojan and steals macOS credentials from developers. The use of an AI-adjacent name in the package demonstrates supply-chain attacks targeting AI developers.

Prompt Injectiontechnology

Show HN: Wolf Defender, a open-weight prompt-injection detection model

Show HN: I built a secure AI mediator to handle my own marital conflicts

PII ExposureChatGPT

How AI Assistants are Moving the Security Goalposts

Krebs on Security reports that AI agents with access to users' files, computers, and online services are blurring security lines between trusted tools and insider threats, as autonomous programs can now automate virtually any task. This highlights the fundamental risk that AI assistants represent for data exfiltration and unauthorized access.

Preventable by Bleep

A DLP proxy could monitor and restrict what data AI agents send to external services, but cannot prevent local machine access if the agent runs client-side.

Insider Threat

Credential ExposuretechnologyClaude

Show HN: Golf Scanner – OSS tool to find and audit every MCP server

Golf Scanner is an open-source tool designed to audit MCP (Model Context Protocol) servers, addressing critical security risks where engineers can expose production databases and sensitive credentials to Claude and other AI tools in seconds. The tool highlights the widespread vulnerability of insecure MCP server configurations that could lead to massive data leakage incidents.

Preventable by Bleep

A DLP proxy could detect and block connections to production databases being configured as MCP servers, but the core vulnerability exists in how MCP servers themselves are set up and validated before data is sent to the AI model.

Show HN: Aegis – Open-source pre-execution firewall for AI agents

Insider Threattechnology

Source Code LeaktechnologyOpenAI Codex SecurityChatGPT

OpenAI Codex Security Scanned 1.2 Million Commits and Found 10,561 High-Severity Issues

OpenAI Codex Security is an AI-powered security agent that scanned 1.2 million commits and found 10,561 high-severity issues. The tool itself addresses security but highlights the risk that code repositories with vulnerabilities are being analyzed by AI systems.

Preventable by Bleep

A DLP proxy could monitor and control what source code is sent to OpenAI's Codex Security service, preventing exposure of proprietary or sensitive code to external AI systems.

Cylake Offers AI-Native Security Without Relying on Cloud Services

Cylake offers an AI-native security platform that analyzes security data locally for organizations concerned about data sovereignty, avoiding cloud-based AI services. This is a product announcement addressing data sovereignty concerns rather than reporting a specific incident.

Securitytechnology

North Korean APTs Use AI to Enhance IT Worker Scams

Credential Exposure

Credential Exposuregovernment

Transparent Tribe Uses AI to Mass-Produce Malware Implants in Campaign Targeting India