The newest AI resources are on the AI page. For a reading order instead of a date order, see the AI learning path.
| Date Added | Link | Excerpt |
|---|---|---|
| 2026-11-06 NEW 2026 | What Is an Agent Harness? How AI Security Tools Work beginner 7 min read | Library for AI agent harnesses, explaining their role in coordinating AI model interactions with tools and environments for security testing. It details how harnesses manage context, tool selection, execution, and feedback loops, using examples like fuzzing SQL injection candidates with Burp Suite and an MCP server. The text emphasizes that harness design significantly impacts an AI's effectiveness, influencing execution, planning, state management, and verification processes critical for automated security engagements. → bishopfox.com |
| 2026-11-06 NEW 2026 | DocsGPT unauthenticated RCE via FAISS pickle deserialization intermediate 2 min read Deser Python RCE | Advisory for GHSA-m448-vp23-j393 detailing an unauthenticated RCE in DocsGPT. Attackers can exploit a flaw in the `/api/upload_index` endpoint to upload malicious FAISS index files. The application's deserialization of these files via `FAISS.load_local` with `allow_dangerous_deserialization=True` leads to arbitrary command execution, bypassing authentication checks even when `AUTH_TYPE` is configured. This critical vulnerability allows full control over the DocsGPT backend. |
| 2026-11-06 NEW 2026 | Does Daybreak Blue cover autonomous pentesting agents, or do you need Red? beginner Bug Bounty | This content discusses a bug bounty researcher's inquiry about Daybreak Blue's coverage, specifically regarding autonomous pentesting agents like Strix. The researcher seeks clarification on whether Daybreak Blue is compatible with or includes such agents, or if a different tool, "Red," is required. The core of the question revolves around the scope and capabilities of Daybreak Blue in the context of automated security testing. |
| 2026-11-06 NEW 2026 | Guardrails for AI Agents: Why I Put Security Checks at the Gateway and in the Code intermediate Supply Chain | This content discusses implementing security checks for AI agents, specifically a supply-chain assistant. The author advocates for placing security measures at two key points: the "gateway" (likely referring to API management like Apigee X) and "in the code" itself, utilizing tools like Model Armor and Google's Agent framework. The core idea is a layered security approach to protect AI agent interactions and operations. |
| 2026-11-06 NEW 2026 | Before Your Security Team Says Yes: What mcp-bastion 1.4.2 and 1.5.0 Ship for a Shadow Pilot beginner Supply Chain | This content discusses the importance of a "shadow pilot" program for new software releases, using mcp-bastion 1.4.2 and 1.5.0 as examples. It emphasizes the need for a thorough evaluation by a security team before widespread deployment. The analogy of a fire alarm installation highlights the potential disruption and the necessity of careful testing to ensure new features don't negatively impact existing operations. The core message is about proactive risk assessment and pilot testing for security software to avoid unforeseen problems. |
| 2026-11-06 NEW 2026 | Harbor Under Audit: Securing an Android App When an Agent Holds the Pen intermediate Mobile | This guide focuses on securing Android apps in the "agentic era," emphasizing agent security skills. It details practical implementation using Kotlin code for technologies like Google's Play Integrity, SMS Retriever, and Advanced Protection. The content aims to equip developers with the knowledge to fortify their applications against emerging threats in an environment where automated agents play a significant role. |
| 2026-11-06 NEW 2026 | Xoscar unauthenticated cloudpickle RCE (CVE-2026-49950) news Deser Python RCE | Writeup of CVE-2026-49950, detailing an unauthenticated RCE vulnerability in Xorbits Xoscar up to version 0.11.2. Attackers can leverage the actor-pool TCP channel by sending crafted data to `SocketChannel.recv()`, bypassing message type validation. This exploits `cloudpickle.loads()` to trigger pickle reduce gadget execution, enabling arbitrary code execution as the server process, often with root privileges on GPU nodes. |
| 2026-11-06 NEW 2026 | AI-Assisted Fuzzing: Generating libFuzzer Harnesses with a Local LLM intermediate 33 min read Fuzzing | Library for generating libFuzzer harnesses using local LLMs. This approach leverages models like Ollama's qwen3.6:35b-a3b to automate the creation of boilerplate code for fuzzing C libraries, streamlining the process and making fuzzing more accessible. The library's workflow demonstrates a complete pipeline: an LLM generates a harness for a target function like `parse_records`, clang compiles it with AddressSanitizer, and execution reveals vulnerabilities such as a stack-buffer-overflow, as detailed in the provided lab example. |
| 2026-11-06 NEW 2026 | Introducing agentic capabilities in OSS-Fuzz beginner 8 min read Fuzzing | Library for automating OSS-Fuzz tasks, introducing agentic capabilities through skills and lightweight infrastructure. This library supports onboarding new projects, improving code coverage, fixing broken builds, and general OSS-Fuzz operations. It includes specialized skills for Go, Java, C/C++, Python, and Rust fuzzing, alongside a general `oss-fuzz-engineer` skill. Helper scripts enable large-scale automation for tasks like expanding projects or repairing builds. |
| 2026-11-06 NEW 2026 | Typed Template Fuzzing (Typos) intermediate 1 min read Fuzzing | Library for Typed Template Fuzzing (TTF), implemented in the tool Typos. TTF utilizes LLMs for preprocessing seeds into typed templates, enabling semantically aware, mutation-based fuzzing without language-specific parsers or further LLM queries during fuzzing. This methodology has been evaluated across 21 input languages, including C, C++, Go, Rust, SystemVerilog, Scala, Java, PyTorch, and Qiskit, reporting 135 bugs, 128 of which were previously unknown, with an average of 96% of mutants being valid. → arxiv.org |
| 2026-11-06 NEW 2026 | Insomnia: an autonomous AI red team that runs a pentest from recon to report beginner Recon | Insomnia is an autonomous AI red team designed to conduct full penetration tests, from reconnaissance to reporting. Unlike manual pentests, which are infrequent and time-intensive, Insomnia automates the process. It handles initial data gathering and scanning, identifying vulnerabilities that a human tester might take days to discover. The AI's goal is to provide continuous, efficient security assessments. |
| 2026-11-06 NEW 2026 | Stanza RCE via pickle deserialization in lemmatizer intermediate 2 min read Deser Python RCE | Library vulnerability impacting stanza, allowing arbitrary code execution via pickle deserialization when loading untrusted lemmatizer checkpoints. The `Trainer.load` function uses an unrestricted `pickle.loads` on gzipped data stored within checkpoint dictionaries, bypassing the `weights_only=True` protection. This vulnerability, confirmed in stanza versions 1.14.0 and later, is reachable through the `stanza.Pipeline` API by providing a malicious lemma model path. The fix involves using the existing `RestrictedUnpickler` during deserialization. |
| 2026-11-06 NEW 2026 | Agenta Jinja2 evaluator SSTI allows RCE intermediate 1 min read Python RCE SSTI | Advisory GHSA-cfr2-mp74-3763 details a Server-Side Template Injection (SSTI) vulnerability in Agenta's API server, allowing authenticated users to achieve RCE (CWE-1336) by crafting malicious Jinja2 templates. When `template_format="jinja2"` is used in AI critique evaluators, the unsandboxed `Template(content).render()` function executes arbitrary Python code on the server, granting full read/write access to the container's filesystem and secrets. This impacts self-hosted or managed Agenta platform deployments, not standalone SDK usage. |
| 2026-11-06 NEW 2026 | LMCache unauthenticated pickle RCE on multiprocess transport intermediate 3 min read Deser Python RCE | Library with an unauthenticated pickle RCE vulnerability in its multiprocess ZMQ transport. Versions 0.3.9 through 0.5.5 are affected, allowing remote code execution as the LMCache process user when the transport binds to a routable address. The vulnerability, identified as GHSA-w5f2-9q9c-ff4h, arises from insecure deserialization of DeviceIPCWrapper messages using `pickle.loads`. A fix is available in version 0.5.6. |
| 2026-11-06 NEW 2026 | CVE-2026-119048: JeeSite AI CMS chat IDOR exposes conversations news API Sec IDOR | A vulnerability (CVE-2026-119048) in JeeSite AI CMS versions 5.11.0-5.19.1 allows authenticated users to access other users' chat conversations. This insecure direct object reference (IDOR) flaw means that with knowledge or by guessing a chat ID, unauthorized access to private messages is possible. The summary focuses on the type of vulnerability, affected software and versions, and the potential impact. → insightcve.com |
| 2026-11-05 NEW 2026 | Google Warns LLMjacking Attacks Are Costing Some Enterprises Up to $1 Million in Cloud Charges news 6 min read Secrets | Analysis of LLMjacking attacks detailing threat actor UNC6993's methods, including misconfigured service accounts, session hijacking pages exploiting active access tokens, and leveraging stolen Personal Access Tokens. These attacks exploit exposed LLM endpoints or compromised API keys to fraudulently consume cloud resources, with cumulative charges reaching $1 million USD. Vulnerabilities like CVE-2026-40217 in LiteLLM are also exploited for privilege escalation. The emergence of LLM Mixers and services like Jet Night highlights a resale market for unauthorized LLM access. → thehackernews.com |
| 2026-11-05 NEW 2026 | 12 Editor Extension Defaults Behind the Trojanized Claude Code Build, and the Fix for Each news Supply Chain | A trojanized version of Anthropic's Claude code editor extension was distributed, running malicious code upon the editor's launch. This compromised build leveraged 12 default settings within editor extensions to execute its harmful payload. The article details each of these default settings and provides the specific configuration change needed to mitigate the vulnerability and secure the editor. |
| 2026-11-05 NEW 2026 | The AI Pentesting wave. Pay for effort vs pay for impact. beginner 8 min read Bug Bounty Talks | Library that analyzes the impact of AI in pentesting and bug bounty programs, arguing that AI tools, while powerful for coverage and brainstorming, do not replace the pay-for-impact model of bug bounty. It explains that AI pentesting often falls into a pay-for-effort category, requiring human validation and triage to distinguish real vulnerabilities from noise, similar to how SAST tools raised the baseline but didn't eliminate the need for deeper analysis. → intigriti.com |
| 2026-11-05 NEW 2026 | 1 AI Subscription made me 100k$ in 2 months beginner Bug Bounty | The author details earning over $100,000 in two months through bug bounty findings on private HackerOne programs. The summary highlights the significant financial success achieved via AI subscriptions and bug bounty work. The content explicitly states a payout of "$100,000". |
| 2026-11-05 NEW 2026 | AI-Powered Bug Bounty: How I Turned Claude Into a Recon & Hunting Partner intermediate Bug Bounty Burp Recon | This article details how a bug bounty hunter integrated Claude, an AI assistant, into their workflow for reconnaissance and vulnerability hunting. By combining Claude with tools like Burp MCP and Chrome DevTools, the author enhanced their bug bounty process. The content focuses on practical application and skill development to leverage AI for more effective security testing. |
| 2026-11-05 NEW 2026 | Google’s AI read the real news on Poland’s border, then invented a 10,000-troop mobilization. news | Google's AI, when processing news about Poland's border, hallucinated and generated a false report about a 10,000-troop mobilization. This incident highlights the potential for AI to misinterpret information and create inaccurate narratives, even when exposed to factual data. The AI's output was a fabrication, not a reflection of actual events. |
| 2026-11-05 NEW 2026 | Why Row-Level Security Is the Safest Place for Authorization in AI-Built Apps intermediate AuthZ | Row-level security (RLS) is presented as the safest approach for authorization in AI-built applications. The core problem highlighted is that AI coding agents, while proficient, can overlook critical security features like tenant filters when generating code. This omission can lead to data breaches where one tenant can access another's data. RLS, by enforcing data access controls at the database level, ensures that even if an AI-generated query misses a tenant filter, the RLS mechanism will still prevent unauthorized data access. This makes RLS a robust safeguard against common AI coding oversights in authorization. |
| 2026-11-04 NEW 2026 | MCP Prompt Injection at Connect: Lab + Fixes intermediate 7 min read API Sec | Library for mitigating MCP prompt injection, featuring four controls: isolation, capping, binding cache keys to server and caller identity, and pinning instruction digests. The library's accompanying lab reproduces the vulnerability, demonstrating how server-controlled instructions can be injected into a model's system prompt, and validates the effectiveness of these defenses in preventing cross-caller leaks and prompt hijacking, particularly when caches use `cacheScope: public`. |
| 2026-11-04 NEW 2026 | We found 4,972 "hello" messages hiding thousands of malicious commands targeting Flowise intermediate | Researchers discovered a significant vulnerability in Flowise, a low-code tool for building LLM applications. They found 4,972 instances of "hello" messages that were actually masking thousands of malicious commands. These commands were designed to target Flowise instances, posing a security risk. |
| 2026-11-04 NEW 2026 | bREACHYing the Perimeter: How One HTTP Request can Hijack your AI Robot intermediate 12 min read Supply Chain | Tool that exploits five CVEs in the Reachy Mini AI robot, enabling unauthenticated remote code execution and privilege escalation. Attackers can hijack the robot's motor control via a single HTTP request to the `reachy-mini-daemon`'s `api/apps/install` endpoint, executing arbitrary Python packages by pointing the `source_kind` to a malicious tarball. This vulnerability, stemming from a lack of authentication, overly permissive CORS, and a non-functional localhost flag, allows attackers on the same network to gain root access. → jfrog.com |
| 2026-11-04 NEW 2026 | AI Findings are Cheap to Find & Costly to Trust beginner 5 min read | Library for analyzing AI-generated vulnerability findings, detailing how AI lowers report generation costs but not verification, leading to an increase in false positives. It highlights the significant human expert time required to validate these AI-produced reports, citing examples like curl's security team and the Linux kernel's principle of human accountability for all findings, regardless of AI assistance. The library aims to help security leaders optimize for confidence and make better risk decisions rather than simply increasing output volume. → bishopfox.com |
| 2026-11-04 NEW 2026 | Architecting the Autonomous Security Agent: Decoupling the Brain from the Blade beginner | This content explores the architectural design of an autonomous security agent, focusing on decoupling its "brain" (decision-making, strategy) from its "blade" (execution, tools). The core idea is to create a modular system where the agent's intelligence is separate from the actual security testing actions. This separation allows for greater flexibility, adaptability, and reusability of components. The goal is to move beyond simple script execution and build a sophisticated agent capable of complex security assessments. |
| 2026-11-04 NEW 2026 | I leaked a Vector Database and Reconstructed the Secret From Numbers Alone intermediate Secrets | The author successfully exploited a vulnerability, leaking a vector database and reconstructing a secret using only numerical data. This achievement was part of their practical learning experience with OWASP Top 10 LLM&Agentic Applications, following their previous work on prompt injection. |
| 2026-11-04 NEW 2026 | How to Use NVIDIA SkillSpector: Scan an AI Agent Skill Before You Install It beginner Supply Chain | NVIDIA SkillSpector is a new tool that allows users to scan AI agent skills before installing them. This helps ensure the safety and integrity of the skills, preventing potential security risks or malicious behavior. By analyzing a skill's code, SkillSpector can identify vulnerabilities or suspicious patterns, giving users peace of mind before integrating new AI functionalities. This proactive approach to AI skill management enhances user trust and security within the NVIDIA ecosystem. |
| 2026-11-04 NEW 2026 | AI Supply Chain Security: Threats, Challenges, and Solutions beginner Supply Chain | AI systems, beyond algorithms, depend on training datasets and open-source libraries. This AI supply chain presents significant security threats, including data poisoning, model evasion, and intellectual property theft. Challenges involve the complexity of these dependencies and the difficulty in ensuring the integrity of every component. Solutions focus on robust data validation, secure coding practices, dependency scanning, and the use of trusted repositories. Proactive security measures throughout the AI development lifecycle are crucial to mitigate risks and build resilient AI systems. |
| 2026-11-04 NEW 2026 | I Let Claude Code Loose on a Project Full of API Keys. It Leaked Them 6 Times Out of 6. intermediate API Sec Secrets | The author discovered that the AI model Claude repeatedly leaked API keys when given a project containing sensitive credentials. Despite multiple attempts, Claude leaked the API keys on all six occasions. This highlights a significant security vulnerability in the AI's ability to handle and protect sensitive information within code. The author shared this repo, implying that developers should be aware of this potential risk. |
| 2026-11-04 NEW 2026 | Will AI Kill the Bug Bounty Industry? beginner Bug Bounty | AI's rapid advancement threatens the bug bounty industry by overwhelming security teams with findings. As AI models become faster and more efficient at identifying vulnerabilities by 2026, the traditional model of manual bug hunting may become obsolete. This shift necessitates adaptation within the industry to avoid collapse. The content focuses on the impending impact of AI on how bugs are discovered and validated, suggesting a significant change is on the horizon for bug bounty programs. |
| 2026-11-04 NEW 2026 | Shanti Se Poocho AI: Potential API Credential Disclosure news API Sec Bug Bounty | A vulnerability in Shanti Se Poocho AI was reported, concerning a potential disclosure of API credentials. Researcher Het Mehta identified this issue and submitted the report on November 3, 2026. The report details a significant security risk where sensitive API credentials could be exposed, impacting the platform's security. |
| 2026-11-03 2026 | China's First Frontier AI Export Controls: Mofcom Announcement No. 48 news | China has announced its first export controls on frontier artificial intelligence (AI) technologies. This move, detailed in Mofcom Announcement No. 48, signifies a significant step in regulating the transfer of advanced AI capabilities. The announcement outlines specific technologies that will be subject to these controls, reflecting a strategic approach to managing the global dissemination of AI advancements and potentially impacting international AI development and trade. |
| 2026-11-03 2026 | Muse AI agent is a commercial surveillance network disguised as an AI assistant news 9 min read | Analysis of Meta's Muse AI agent reveals it functions as a commercial surveillance network disguised as an assistant, accessing private messages, financial accounts, and personal relationships. Internal documents show Muse analyzes user behavior to influence individuals and uses interactions for AI training by default, despite internal warnings about security and privacy failures. The agent can profile people who have never used the service and has exhibited unauthorized access to private data, including home addresses and sensitive files, with significant vulnerabilities and security flaws identified prior to its accelerated launch. |
| 2026-11-03 2026 | Backporting security fixes at scale with Codex intermediate 5 min read Supply Chain | Tool for accelerating the backporting of security fixes for stable Linux releases, utilizing Codex and GPT-5.6. This application automates tasks like identifying missing fixes, adapting patches, running tests, and preparing submission packets, significantly reducing the human effort required to address vulnerabilities like CVE-2026-98432 in software such as Ubuntu LTS, Fedora, and the Linux kernel. The project has successfully delivered fixes for 149 distinct CVEs to stable Linux kernel releases and numerous packages to Fedora and Ubuntu. → blog.trailofbits.com |
| 2026-11-03 2026 | The Fireworks AI Breach: How Stolen Internal Credentials Exposed Customer Secrets news Secrets | An attacker, operating under the moniker "Black October 2026," exploited stolen internal Fireworks AI credentials to access and read sensitive environment variables. This breach exposed customer secrets. The article is the first part of a series detailing the incident. |
| 2026-11-03 2026 | Bug bounty hunters: Have you found security bugs in AI-generated web or mobile apps? beginner AuthZ Bug Bounty Mobile | This post explores the security implications of AI-generated web and mobile applications. The author created an iPhone app using Codex with no prior Swift experience, prompting an investigation into potential vulnerabilities. The content asks bug bounty hunters if they have discovered security bugs in AI-generated apps and invites discussion on the topic. |
| 2026-11-03 2026 | The Tensorlake npm Compromise: How the Shai-Hulud Worm Jumped Into an AI SDK news Supply Chain | In a sophisticated supply chain attack, version 0.5.144 of the Tensorlake npm package was compromised by the "Shai-Hulud" worm. A single installation of this version was sufficient to deploy the malicious worm, which was designed to steal credentials. This incident highlights a significant vulnerability within AI SDKs distributed through npm, demonstrating how quickly malicious code can infiltrate critical development tools. |
| 2026-11-03 2026 | The Claude Code VS Code extension backdoor on the Cursor Marketplace should change how you treat… news Supply Chain | A backdoor was discovered in the Claude Code VS Code extension, available on the Cursor Marketplace. This incident highlights the critical need for greater caution when updating developer tools. The article emphasizes practical measures organizations can implement to prevent editor extensions from gaining excessive control over their systems, urging a reevaluation of how such tools are managed and secured. |
| 2026-11-03 2026 | The OpenCode Console Breach: Why Your AI Provider Keys Are Now a Target news | Attackers breached OpenCode's legacy Console system, gaining access to sensitive data including AI provider API keys. This incident, part of "Black October 2026," highlights the growing risk to AI infrastructure. The compromise means these stolen keys could be misused to access or exploit AI services provided by third parties, posing a significant threat to organizations relying on those services. |
| 2026-11-03 2026 | Malicious Claude Code Extension Served in Cursor Supply Chain Attack news 2 min read Recon Supply Chain | Extension alert detailing a supply chain attack on Cursor users via malicious versions (2.1.312, 2.1.313) of Anthropic's Claude Code extension. The compromised extension contained JavaScript for reverse SSH tunnels and host reconnaissance, including OS commands and Git email collection. While admin rights prevented full SSH tunneling, standalone persistence scripts could remain after uninstallation. Cursor released a detection script and IOCs to aid users in identifying and remediating potential infections. → securityweek.com |
| 2026-11-02 2026 | Northflank security breach leads to abuse of customers' AI API keys news 3 min read API Sec Secrets | Writeup of the Northflank security breach, where attackers gained access to customer secrets, leading to the unauthorized abuse of AI API keys. This incident, part of a wider "Black October" trend targeting AI developer platforms like Railway, OpenCode, and Warp, highlights the risks of LLMJacking. Attackers exploited stolen credentials to access AI models, generating significant API charges or reselling access, as evidenced by millions of Amazon Bedrock API calls reported by Wiz. Northflank responded by advising credential rotation and enabling audit logs. → bleepingcomputer.com |
| 2026-11-02 2026 | One Forgotten Key, One Poisoned Release: What the LiteLLM Supply-Chain Attack Means for Your… news Python Supply Chain | The LiteLLM supply-chain attack highlights a critical security vulnerability. A forgotten, exposed API key allowed an attacker to inject malicious code into a popular open-source library. This compromised version was then released, potentially affecting numerous downstream users who relied on LiteLLM. The incident underscores the risks associated with software supply chains and the importance of rigorous security practices, including key management and code auditing, to prevent similar attacks. |
| 2026-11-02 2026 | A Backdoored Claude Code Extension Reached Thousands of Developers news Supply Chain | A malicious Claude code extension was distributed through the VS Code Marketplace, potentially impacting thousands of developers. The extension contained a backdoor, raising significant security concerns for users who installed it. The exact number of developers affected and the extent of any compromise are still under investigation. |
| 2026-11-02 2026 | Your AI Coding Extension Is Now Part of Your Attack Surface beginner Supply Chain | AI coding extensions, while boosting developer productivity, introduce significant new attack vectors. These tools, often integrated directly into IDEs, can become part of an organization's attack surface. Malicious actors could exploit vulnerabilities in these extensions to gain access, exfiltrate sensitive code, or even inject malicious code into the development pipeline. Organizations must therefore treat AI coding extensions with the same security scrutiny as other software components, focusing on secure configuration, vulnerability management, and access controls to mitigate associated risks. |
| 2026-11-02 2026 | How I Approach Prompt Injection Testing: A Practical Guide to AI Security | by Samadhan Shimple intermediate | This article by Samadhan Shimple outlines a practical approach to prompt injection testing in AI security. It focuses on identifying vulnerabilities where malicious prompts can manipulate AI models into unintended behaviors. The guide likely details methodologies and techniques for discovering these weaknesses. The content aims to help security professionals and developers better understand and mitigate prompt injection risks to enhance AI system security. |
| 2026-11-02 2026 | ZDI-26-797 Meta FAIRChem pickle deserialization news Deser Python | This advisory details a pickle deserialization vulnerability in Meta FAIRChem, assigned identifier ZDI-26-797. Successful exploitation of this bug allows an attacker to achieve remote code execution. The vulnerability arises from the improper handling of serialized data. → zerodayinitiative.com |
| 2026-11-02 2026 | ZDI-26-845 MLflow deserialization of untrusted data news Deser Python RCE | Writeup of ZDI-26-845, detailing a remote code execution vulnerability in MLflow's `_load_model` function. This flaw stems from improper validation of user-supplied data, leading to the deserialization of untrusted input. Exploitation requires user interaction, such as visiting a malicious page or opening a crafted file, enabling attackers to execute arbitrary code within the context of the current user. The vulnerability is patched in MLflow 3.15.0. → zerodayinitiative.com |
| 2026-11-02 2026 | MLflow unauthenticated full-read SSRF in webhook delivery intermediate 3 min read Python SSRF | Writeup of CVE-2026-64849 in MLflow details an unauthenticated full-read Server-Side Request Forgery (SSRF) vulnerability within the webhook delivery mechanism. This vulnerability, also documented as GHSA-7gwp-5pfp-969j, arises from bypassed validation of HTTP redirects and DNS rebinding, allowing attackers to access internal resources. The advisory notes that MLflow versions between 3.3.0 and 3.14.0 are affected, with a vendor fix available in version 3.15.0. |