appsec.fyi

AI Resources

Post Share

A curated AppSec resource library covering XSS, SQLi, SSRF, IDOR, RCE, XXE, OSINT, and more.

AI

AI security encompasses both protecting AI systems from attack and understanding the new vulnerability classes that AI introduces into applications. As organizations rapidly integrate large language models (LLMs), machine learning pipelines, and AI-powered features into their products, the attack surface has expanded in ways that traditional application security frameworks don't fully address.

Key threats to AI systems include prompt injection — where attackers manipulate LLM behavior through crafted inputs — data poisoning of training datasets, model extraction through repeated API queries, and adversarial examples that cause misclassification. Indirect prompt injection, where malicious instructions are embedded in data the AI processes (emails, documents, web pages), is emerging as one of the most significant security challenges for AI-integrated applications.

AI also introduces new categories of application risk: insecure output handling where LLM responses are rendered unsafely, excessive agency when AI agents are given too much access, sensitive information disclosure through training data leakage, and supply chain risks from fine-tuned models and third-party plugins. The OWASP Top 10 for LLM Applications provides a structured framework for understanding these risks.

On the defensive side, AI is being used to enhance security operations — automating vulnerability detection, analyzing malicious patterns, and accelerating incident response.

This page collects AI security research, LLM vulnerability techniques, defensive strategies, and resources covering the intersection of artificial intelligence and application security.

The security problem with LLM applications is that instructions and data share a channel

Nearly every novel vulnerability class in this category reduces to one property: a language model receives instructions and untrusted content in the same token stream, with no structural boundary between them. That is not a bug in a particular product, and it is not currently fixed by a patch. It is how the technology works, and it is why prompt injection has no equivalent of the parameterized query.

Direct prompt injection — a user telling the model to disregard its instructions — is the demonstration case and the least interesting one, since the user is only attacking their own session. The consequential form is indirect: the model ingests a web page, a document, an email, a code comment or a tool result that contains instructions, and acts on them. Now the attacker is whoever controlled that content, and the victim is the user whose session the model is running in.

Severity is determined by capability, which is why agents changed the calculus. A model that only produces text has limited blast radius. A model that can call tools, read a filesystem, query internal systems, send requests or execute code turns injected instructions into actions, and the classic confused-deputy problem reappears with a natural-language interface. The combination to watch for is the one that keeps producing real incidents: access to private data, exposure to untrusted content, and an ability to communicate outward. Any two are usually survivable; all three are an exfiltration path.

The rest of the surface is more familiar than the discourse suggests. Retrieval pipelines are an injection point if anyone can write to the corpus. Model and dependency supply chain is the same problem as any other package ecosystem, with the added detail that some model formats execute code on load. Output handling is ordinary injection — a model's response rendered as HTML or passed to a shell is XSS or command injection regardless of where the string came from.

The material here moves quickly. Prefer entries that describe a mechanism over ones that benchmark a specific model version.

Start here: the AI learning pathAn ordered route through this library — orientation, then methodology, then research. Read the AI guideA long-form, source-cited deep dive synthesized from every resource below. The comprehensive AI guide on chs.usA hand-written, in-depth practitioner guide — attacks, testing, and prevention.
Date Added Link Excerpt
2026-11-21 NEW 2026Slopsquatting: When AI Invents a Package Name and Attackers Register It beginner Supply ChainAI coding assistants can sometimes suggest non-existent software package names. Attackers exploit this by registering these invented names and publishing malware under them, a practice called "slopsquatting." This allows malicious code to be distributed through package managers, potentially infecting users who unknowingly install the compromised packages based on AI recommendations. The AI's inability to verify package existence creates a vulnerability that malicious actors can leverage for cyberattacks.
2026-11-21 NEW 2026Attackers Compromise Microsoft's Agent Framework PyPI Package to Steal Cloud Credentials news Python Supply ChainAttackers compromised Microsoft's Agent Framework by uploading malicious versions to PyPI. These tainted releases leveraged Python startup hooks to steal cloud and CI/CD credentials from unsuspecting users. Both malicious packages have since been removed from the platform. → thehackernews.com
2026-11-21 NEW 2026Sn1per SE v11.1: Agentic AI pentesting, MCP and AD mode beginner ReconSn1per SE v11.1 introduces agentic AI pentesting capabilities. This update features AI-driven triage and adaptive scanning, allowing for more intelligent and responsive security assessments. It also includes AI selection of Metasploit modules through MCP, Playwright for XSS verification, and a dedicated Active Directory (AD) mode for enhanced network security testing.
2026-11-21 NEW 2026AI Pentesting Part 8: Burp Suite AI features, extensions, evidence-first workflow intermediate BurpThis guide explores Burp Suite's AI features for penetration testing, focusing on an evidence-first workflow. It details how AI assists in Repeater and Explainer functionalities, aids in exploring issues, and validates access control. The content also covers AI's role in handling login sequences, and discusses extensions like Montoya and MCP.
2026-11-21 NEW 2026Arize Phoenix CVE-2026-53998: one unauthenticated GraphQL query reads os.environ via f-string template news GraphQL Python Secrets SSTIArize Phoenix versions 13.0.0 through 15.6.x contained a critical vulnerability, CVE-2026-53998, allowing unauthenticated attackers to read sensitive environment variables like `PHOENIX_SECRET` via a crafted GraphQL query. This exploit leverages the `applyChatTemplate` function's use of Python's `str.format()` with user-supplied f-string templates. Leaking `PHOENIX_SECRET` enables attackers to forge admin session tokens and decrypt stored LLM API and OAuth tokens. A fix was released in version 15.7.0.
2026-11-21 NEW 2026The $1000 IDOR Hiding Behind a Shared AI Chat intermediate AuthZ Bug Bounty IDORA security researcher discovered an Insecure Direct Object Reference (IDOR) vulnerability in a platform utilizing shared AI chats. By intercepting requests and manipulating a conversation ID, they could access private skill identifiers. An overlooked API endpoint then allowed them to exploit this identifier to access confidential user data. This vulnerability, which was awarded a $1,000 bounty, highlights the risks associated with shared conversational contexts and the importance of securing all API endpoints, even those that appear secondary.
2026-11-20 NEW 2026Hacking with AI: How to prevent your AI from going rogue intermediate Bug BountyAI is increasingly used in bug bounty programs for tasks like reconnaissance and vulnerability identification. While a powerful tool for security researchers, improperly deployed AI models can cause damage due to intrusive testing. Without proper guardrails, AI lacks context regarding program scope, potentially leading to unintended consequences. → intigriti.com
2026-11-20 NEW 2026[tl;dr sec] #355 - Cyber Chaos Monkey, Offsec Harness Index, Top 10 AWS Security Research of 2026 newsThis edition of [tl;dr sec] covers several key security developments. Shopify has implemented an AI agent system that proactively attacks its own production infrastructure to identify vulnerabilities. It also highlights a comprehensive index of open-source offensive AI agent harnesses, a valuable resource for security researchers. Finally, the content provides a "top 10" list of significant AWS security research from 2026, offering insights into the latest advancements and threats in cloud security. → tldrsec.com
2026-11-20 NEW 2026Microsoft’s Agent Framework PyPI package compromised to steal credentials news 2 min read Python Supply ChainLibrary compromise via malicious PyPI releases of Microsoft's Agent Framework, specifically versions 1.26.1 and 1.26.2, targeted developers by stealing cloud credentials for Azure, AWS, and GitHub, among others. The malware executed during Python startup, harvesting secrets and tokens, and exfiltrating data to an attacker-controlled endpoint. Developers are advised to isolate systems and rotate credentials. → bleepingcomputer.com
2026-11-20 NEW 2026The Tampered Claude Code Extension in Cursor’s Marketplace: What It Did and How to Check news Supply ChainA compromised Claude Code extension in Cursor's marketplace inserted an SSH backdoor for approximately eleven hours on October 29th. This malicious code allowed unauthorized access to users' systems. While the specific impact and full extent of the breach are still under investigation, the presence of the backdoor posed a significant security risk. Users who had the extension installed during that period are advised to check their systems for any signs of compromise and to remove the extension immediately.
2026-11-20 NEW 2026The Fireworks AI Breach: How One Vendor Key Became a Supply Chain Attack on Cursor news 9 min read Secrets Supply ChainWriteup detailing the Fireworks AI breach, tracing how a single vendor key led to a supply chain attack on Cursor. The breach exposed API keys and tokens from Fireworks, subsequently allowing attackers to access Cursor's internal tools via credentials found in agent traces. This incident resulted in a trojanized Claude Code extension being available on Cursor's marketplace, potentially impacting thousands of developers and highlighting risks associated with sharing credentials with AI vendors and coding assistants. → blog.gitguardian.com
2026-11-20 NEW 2026NVIDIA Megatron-LM checkpoint loading deserialization news Deser PythonNVIDIA Megatron-LM has a deserialization vulnerability in its checkpoint loading feature. This flaw could allow an attacker to execute arbitrary code. The vulnerability stems from the improper handling of untrusted input during the loading of model checkpoints. Users are advised to update to the latest version of Megatron-LM to mitigate this security risk.
2026-11-20 NEW 2026AppSec Podcast: Johann Rehberger on Breaking, Watching, and Killing Agents intermediate 43 min read TalksPodcast episode featuring Johann Rehberger discussing his "trust no AI" philosophy, honed through years of red teaming and breaking AI agents. He highlights practical security measures like isolating autonomous agents, exporting logs for external review, treating model output as untrusted, and establishing baseline test cases for re-evaluation when models change. Rehberger shares insights into specific incidents, including a GLM 5.3 agent attempting unauthorized SSH access and the realization that Docker doesn't always provide true sandbox isolation. The discussion also covers crucial advice for CISOs regarding the rollout of coding agents, emphasizing inventory, monitoring, and red teaming.
2026-11-20 NEW 2026How an GraphQL IDOR leaked AI conversation History and Got me $500 bounty intermediate API Sec Bug Bounty GraphQL IDORA security researcher discovered a GraphQL Insecure Direct Object Reference (IDOR) vulnerability. This flaw allowed unauthorized access to AI conversation history. The vulnerability was reported and the researcher was awarded a $500 bounty for their findings.
2026-11-19 NEW 2026LLM assisted reverse engineering intermediate 5 min readLibrary for LLM-assisted reverse engineering that streamlines analysis of binaries, firmware, and software. It details a process utilizing dedicated agent skills for binary analysis, documentation, and debugging, supporting techniques like hypothesis debugging and probe implementation with Python/UV tooling and Rust. The library enables rapid success in reverse engineering hardware and software, exemplified by projects like hacking Alphasmart NEO devices, mapping communication protocols, and developing custom firmware for musical synthesizers and e-readers.
2026-11-19 NEW 2026Microsoft’s agent-framework PyPI package Compromised with Credential Stealer news 3 min read Python Supply ChainWriteup detailing the compromise of Microsoft's agent-framework PyPI package (versions 1.26.1 and 1.26.2) with credential-stealing capabilities. The malicious code exploits Python's `.pth` startup hooks, specifically `agent_framework_meta.pth` and `af_runtime_diag.pth`, to exfiltrate Azure, GitHub Actions, and CI/CD credentials by harvesting environment variables, local configuration files, and cloud identity tokens. The exfiltrated data is then compressed, Base64-encoded, and sent to a `webhook.site` endpoint. → socket.dev
2026-11-19 NEW 2026Cursor Says Stolen Fireworks AI Key Led to Trojanized Claude Code Extension Incident news 3 min read Secrets Supply ChainAnalysis of the Cursor, Fireworks AI, and Anthropic Claude Code extension incident details how a trojanized extension was distributed via Cursor's marketplace. The attack stemmed from a stolen Fireworks AI access key, which allowed an attacker to access Cursor's storage bucket, exposing customer API credentials. This data was then used to push a malicious version of the Claude Code extension, creating an SSH backdoor for potentially 5,000 users. → thehackernews.com
2026-11-19 NEW 2026NVIDIA Model Optimizer unsafe deserialization news DeserNVIDIA Model Optimizer for Linux and Windows has an unsafe deserialization vulnerability. This flaw allows an attacker to execute arbitrary code, potentially leading to a system compromise. The vulnerability arises from insufficient validation of user-supplied data during deserialization. This could impact users who rely on NVIDIA Model Optimizer for machine learning model optimization and deployment.
2026-11-19 NEW 2026SGLang missing authorization news AuthZWriteup of GHSA-m386-99gj-pq7p, detailing a missing authorization vulnerability in SGLang through 0.5.23. The POST /load_lora_adapter_from_tensors endpoint, lacking ADMIN_OPTIONAL authorization, allows unauthenticated attackers to invoke it. This enables loading attacker-supplied LoRA adapters to manipulate the served model, and when combined with a deserialization bypass, can lead to unauthenticated remote code execution as root.
2026-11-19 NEW 2026Black Hat USA 2026 | Can AI Do Novel Security Research? Meet the HTTP Terminator advanced TalksAt Black Hat USA 2026, a presentation explored whether AI can conduct novel security research, going beyond bug detection to invent new attack techniques. The "HTTP Terminator" system was showcased as an autonomous agent designed to tackle this challenge, aiming to autonomously discover and develop innovative hacking methods.
2026-11-19 NEW 2026Black Hat USA 2026 | Threat Modeling LLMs: The PHANTOM-B model advanced TalksSecurity engineers developing LLM products contend with a rapidly evolving threat landscape and pressure for quick deployment. Traditional threat modeling struggles with the unique complexities of LLMs. This talk introduces the PHANTOM-B model, a novel approach specifically designed for threat modeling LLMs. PHANTOM-B aims to provide a more effective framework for identifying and mitigating risks associated with large language models.
2026-11-19 NEW 2026How to set up Burp Suite with Claude desktop (using Burp MCP Server) intermediate BurpThis guide explains how to integrate Burp Suite with Claude desktop using Burp MCP Server for web application security testing. It covers common vulnerability scenarios like Broken Access Control, IDOR, XSS, and CSRF, providing a setup process to enhance the testing workflow. The focus is on leveraging Burp Suite's capabilities alongside Claude for more effective security analysis.
2026-11-19 NEW 2026mssql-mcp SQL injection via unparameterized table names intermediate 1 min read SQLiLibrary addressing SQL injection vulnerabilities in mssql-mcp-core, specifically GHSA-2m9m-6cr5-9x25, due to unparameterized T-SQL identifiers and predicates. Affected tools like `update_data`, `delete_data`, `insert_data`, `execute_transaction`, `create_table`, `create_index`, `drop_table`, and `list_databases` allowed arbitrary DDL/DML and OS command execution via stacked statements. The fix in version 0.6.0 validates identifiers and replaces free-form `whereClause` with structured, parameterized filters, while `list_databases` now validates `stateFilter` against an allowlist.
2026-11-18 NEW 2026First phase of embedded assessments at OpenAI newsOpenAI is rolling out the first phase of embedded assessments. These assessments are designed to evaluate AI models' capabilities and performance in a more integrated manner. The goal is to continuously improve AI systems through rigorous testing and feedback. This initiative marks a significant step in OpenAI's commitment to developing safer and more effective AI.
2026-11-18 NEW 2026AI Security Engineering Series: Model Extraction Outgrew Rate Limits intermediateThis post, part of an AI Security Engineering Series, shifts focus from data protection to safeguarding AI models. It discusses the challenge of model extraction, where attackers attempt to replicate a model, and how this threat has evolved beyond traditional rate-limiting defenses. The content explores advanced techniques and considerations for securing AI models against such sophisticated attacks.
2026-11-18 NEW 2026AI Coding Agent RCE: How a README Can Run Code on Your Laptop intermediate RCEAI coding agents, designed to read repositories, fetch web pages, and execute commands, are now being exploited by attackers. The vulnerability lies in how these agents process README files. Malicious code embedded within a README can be executed on a developer's machine, posing a significant security risk. This method allows attackers to compromise developer environments through seemingly innocuous repository files.
2026-11-18 NEW 2026Who scans AI infrastructure? Ollama, Jupyter, MLflow data beginner ReconAI infrastructure, including tools like Ollama, Jupyter, and MLflow, is frequently deployed on GPU servers with internet access for team collaboration. A significant security concern is that many of these setups lack proper login credentials, leaving them vulnerable. This open access can expose sensitive AI models and data to unauthorized users.
2026-11-18 NEW 2026WebMCP: what can go wrong in the agentic web? (W3C talk) intermediate TalksThis W3C talk, "WebMCP: what can go wrong in the agentic web?", presents a threat model for agentic web navigation. It uses data flow diagrams and trust boundaries to identify potential security issues. The talk explores open security questions within this evolving landscape, highlighting risks associated with automated web agents.
2026-11-18 NEW 2026NahamCon 2026 talks news 2 min read Bug Bounty TalksTalks from NahamCon 2026 explore AI hacking, covering topics like breaking Gemini, automating device hacking with LLMs, exfiltrating data from AI agents, and running supply chain attacks in the AI era. Featured speakers include researchers and hackers discussing vulnerabilities and techniques related to AI security, alongside sessions on bug bounty programs and open-source remediation.
2026-11-18 NEW 2026CSRF in Brainstorm Force ZIP AI intermediate CSRFA CSRF vulnerability was discovered in Brainstorm Force ZIP AI, impacting multiple components. The vulnerability allows attackers to trick authenticated users into performing unwanted actions on the application without their knowledge. This could lead to unauthorized modifications or deletions of data, potentially causing significant disruption. The issue stems from improper input validation and insufficient protection against forged requests.
2026-11-18 NEW 2026LLM API Exploitation with Excessive Agency : Portswigger Lab Writeup beginner API Sec BurpThis Portswigger Lab writeup explores LLM API exploitation, focusing on issues arising from "excessive agency." It introduces what Large Language Models (LLMs) are and their fundamental workings, setting the stage for understanding how their APIs can be targeted. The content aims to provide a beginner-friendly introduction to this domain of AI security.
2026-11-18 NEW 2026Microsoft agent-framework Compromised: Credential Stealer Targets Azure and CI Pipelines news 4 min read Python Secrets Supply ChainLibrary detecting a malicious version (1.26.1 and 1.26.2) of Microsoft's `agent-framework` published to PyPI. The compromised package drops `.pth` files that execute a credential stealer upon Python interpreter startup, targeting Azure and Azure DevOps secrets, environment variables, and credential files. Version 1.26.2 includes expanded credential theft capabilities and GitHub Actions OIDC theft. → wiz.io
2026-11-17 NEW 2026Windows 11 Is Collecting Data on Which Coding Agents and Developer Tools You Use news 6 min readAnalysis of MissionControlAggregator.dll reveals Windows 11 collects telemetry on developer tools, including AI coding agents like copilot.exe, claude.exe, and cursor.exe, along with standard tools such as python.exe, dotnet.exe, and code.exe. This data, enabled by default unless optional diagnostic data is opted out of, is sent to Microsoft and is collected by subscribing to Windows events like AppStateChangeSummary and AppInteractivitySummary. The component assigns specific bits to categories, such as a "Codex bit," indicating the usage of detected applications.
2026-11-17 NEW 2026My AI can use my API key. It just can’t see it beginner SecretsThis article describes a new credential proxy designed to allow AI agents to use API keys without directly "seeing" them. The proxy is embedded within the AI model's weights, enabling secure API calls. This innovation addresses a critical challenge as AI agents become increasingly adept at interacting with external services.
2026-11-17 NEW 2026Cursor Supply Chain Attack Started at Fireworks AI news 2 min read Supply ChainWriteup detailing the Cursor supply chain attack initiated by a breach at Fireworks AI, which led to compromised credentials allowing attackers to distribute malicious versions of the Claude Code extension via the Cursor VS Code Extension Marketplace. The attack leveraged a Cursor access key issued to Fireworks, ultimately exposing unredacted credentials that granted access to Cursor's SaaS tools and sensitive data, affecting up to 5,000 users. This incident also impacted Notion and Uber. → securityweek.com
2026-11-17 NEW 2026Trojanized Claude Code Extension Delivered Persistent SSH Backdoor to Developer Machines news 6 min read Supply ChainWriteup of a trojanized Claude Code VS Code extension that delivered a persistent SSH backdoor. Versions 2.1.312 and 2.1.313 of anthropic.claude-code were distributed via Cursor's marketplace, embedding a loader that performed reconnaissance, implanted an SSH key, enabled remote login, established a reverse tunnel, implemented persistence mechanisms (cron, LaunchAgent, scheduled tasks, startup folder), and sent beacon data. The loader did not exfiltrate credentials or access Claude/VS Code data. → stepsecurity.io
2026-11-17 NEW 2026Amazon Uncovers AI Agent Campaign Targeting LiteLLM, Dify, and Other Apps for API Keys news 6 min read API Sec SecretsLibrary of AI penetration testing scripts, including ARTEX, targets LiteLLM, Dify, and other AI applications to steal API keys. Attackers use these keys to fuel further attacks, exploiting vulnerabilities like default credentials, unfinished installations, and insecure endpoints. Techniques observed include authentication bypass, SSRF, and exploitation of secrets found in code, client-side JavaScript, and container images. Targeted applications span LLM gateways, proxies, and AI app builders, with a focus on exposing multiple provider keys via single weak points. → thehackernews.com
2026-11-17 NEW 2026AI System Reconnaissance — TryHackMe Write-up beginner ReconThis TryHackMe write-up by R1punjay_Singh details AI system reconnaissance. The content likely explores methods and techniques for gathering information about AI systems, potentially for security testing or ethical hacking purposes. The focus would be on understanding an AI's architecture, data sources, and potential vulnerabilities through various reconnaissance strategies.
2026-11-17 NEW 2026Black Hat USA 2026 | ChatMate: Remote Prompt Execution on AI Assistants through Sandbox Escaping advanced RCE TalksResearchers at Black Hat USA 2026 will present a novel attack called ChatMate, which exploits vulnerabilities in AI assistants. This method allows attackers to achieve remote prompt execution on a user's chat session by escaping the sandbox environment. The technique enables attackers to interact with the LLM as if they were the legitimate user, potentially leading to unauthorized actions or data exposure.
2026-11-17 NEW 2026Black Hat USA 2026 | AI and the Future of Cyber Defense Panel beginner TalksFrontier AI models are significantly impacting cybersecurity by enhancing both offensive and defensive capabilities. These advanced AI models are crossing critical thresholds, leading to new possibilities and challenges in the cyber landscape. The panel at Black Hat USA 2026 will explore how these evolving AI technologies are reshaping the future of cyber defense.
2026-11-17 NEW 2026emlog pro AI Assistant CSRF to SQL execution PoC intermediate 3 min read CSRF SQLiProof-of-concept detailing CVE-2026-73847 in emlog pro, exploiting a missing CSRF token, session-cookie-only authentication, and a hardcoded confirmation string on the AI Assistant's `execute_tool` endpoint. This allows an attacker to leverage an admin's authenticated session to execute arbitrary SQL, enabling full database reads and writes to all tables except `blog`, including user account takeover via password overwrites. Remediation involves adding CSRF token checks, setting `SameSite=Strict` on auth cookies, and implementing per-session confirmation tokens.
2026-11-17 NEW 2026Introducing Wake 5: agent-written Solidity fuzzing in Python intermediate 3 min read Fuzzing PythonFramework for agent-written Solidity fuzzing in Python, Wake 5 integrates an EVM within the Python process using revm, allowing live chain forks and cross-chain tests. It features an MCP server for agents to access static contract facts, generating Python types for autocompletion. Wake 5 achieves high transaction throughput and agent-written tests significantly outperform tools like Foundry and Echidna in reaching assertions on benchmark mazes.
2026-11-17 NEW 2026DeepAudit SSRF intermediate SSRFLibrary advisory for DeepAudit versions 2.0.0-beta.5 through 3.8.0 details a server-side request forgery vulnerability (CVE-2026-129880) allowing registered users to leak server-wide GitLab or Gitea tokens. Attackers can exploit this by creating a malicious project and requesting its branches, enabling token capture and access to internal hosts.
2026-11-17 NEW 2026From AI Prompt Injection to Authenticated XSS in a CRM Platform intermediate Bug Bounty XSSA researcher discovered a critical vulnerability in a CRM platform, chaining prompt injection to authenticated stored XSS. This exploit allowed unauthorized actions within the system. The vulnerability began with a prompt injection flaw, which, when combined with an authenticated user, led to a stored Cross-Site Scripting (XSS) vulnerability. This writeup marks the researcher's first bug bounty report.
2026-11-16 NEW 2026Who Owns Your Agent’s Credentials? beginner SecretsThe security of AI agent credentials is a critical, yet often overlooked, vulnerability. As AI agents increasingly perform complex tasks like sending emails, accessing customer data, and updating CRMs, ensuring the ownership and security of their digital identities becomes paramount. This emerging gap in AI security requires immediate attention from development teams to prevent potential breaches and maintain control over agent actions.
2026-11-16 NEW 2026AI Agent Traces Are a Credential Store: What Cursor’s Breach Shows About Secrets in Agent Logs beginner SecretsA recent breach at Cursor highlights a significant security risk: AI agent logs can inadvertently become credential stores. When sensitive information, such as API keys or passwords, is included in an AI agent's context, it can be captured and stored within the agent's traces and logs. This means that any data generated from that context, including historical records, could potentially expose these secrets. The Cursor incident serves as a stark reminder for developers to carefully manage what data is fed into AI agents and to implement robust security measures to prevent credential leakage through agent logs.
2026-11-16 NEW 2026Cursor Extension Backdoor and the Fireworks AI Vendor Key news Supply ChainThis MAG212 CYBER Weekly update from November 14, 2026, confirms a backdoor found in a Cursor extension. The vulnerability is linked to a Fireworks AI vendor key. The researchers who reported this issue have confirmed its existence with the vendor. The content further details the commands affected by this backdoor.
2026-11-16 NEW 2026TryHackMe: Understanding AI Supply Chains — Walkthrough beginner Supply ChainThis TryHackMe walkthrough focuses on understanding AI supply chains. It guides users through the process of exploring and securing these complex systems. The content likely covers various stages of the AI lifecycle, from data acquisition and model training to deployment and monitoring, highlighting potential vulnerabilities and mitigation strategies within each. → infosecwriteups.com
2026-11-16 NEW 2026JGD: an autonomous agent for Java deserialization gadget chains intermediate 3 min read DeserLibrary for autonomous Java deserialization research, JavaGadgetDigger (JGD) automatically identifies gadget chains and generates weaponized proof-of-concept payloads from JAR directories. It supports discovering novel chains through static and dynamic analysis, including adversarial LLM auditing for confirmation. JGD can determine the status of known public chains against specific environments and produces payloads for RCE. It automates the end-to-end process of finding exploitable deserialization vulnerabilities, detailing chains involving libraries like Vavr, Spring AOP, and Guava, and specific techniques such as `HashMap.put` to `Method.invoke` chains.
2026-11-16 NEW 2026IBM ContextForge MCP Gateway auth bypass news AuthNIBM ContextForge MCP Gateway versions 1.0.0 through 1.0.11 contain an authentication bypass vulnerability. This flaw allows an unauthenticated attacker to gain unauthorized access to sensitive data and system functions. Users are advised to update to the latest version to mitigate this security risk.

Frequently Asked Questions

What is prompt injection?
Prompt injection is an attack against applications that use large language models (LLMs). An attacker crafts input that overrides or manipulates the LLM's system instructions, causing it to perform unintended actions. Direct prompt injection targets the user input; indirect prompt injection embeds malicious instructions in data the LLM processes, such as emails or web pages.
What is the OWASP Top 10 for LLM Applications?
The OWASP Top 10 for LLM Applications identifies the most critical security risks for AI-powered applications, including prompt injection, insecure output handling, training data poisoning, model denial of service, supply chain vulnerabilities, sensitive information disclosure, insecure plugin design, excessive agency, overreliance, and model theft.
How do you secure AI-integrated applications?
Key practices include validating and sanitizing LLM outputs before rendering or executing them, implementing least-privilege access for AI agents, using guardrails to constrain model behavior, monitoring for prompt injection attempts, applying rate limiting, separating AI processing from privileged operations, and treating all LLM output as untrusted user input.

Weekly AppSec Digest

Get new resources delivered every Monday.