AI Sec Watch

Crawl4AI's Docker API had a security flaw in its SSRF protection (a filter meant to block requests to internal servers). An attacker could bypass this filter by encoding internal IP addresses using IPv6 transition forms like NAT64, 6to4, or IPv4-mapped addresses, potentially accessing internal services and cloud metadata endpoints without needing credentials.

This is a subscriber-only eBook collection from MIT Technology Review featuring six stories about how military organizations are using AI models to help make decisions. The stories were originally published between April 2025 and April 2026 and have been updated to reflect recent developments in military AI applications.

Crawl4AI, a Docker API server for web crawling, had seven critical security vulnerabilities including arbitrary file writes, SSRF (server-side request forgery, where attackers trick the server into making requests to internal networks), authentication bypass on monitoring endpoints, stored XSS (cross-site scripting, where malicious code executes in users' browsers), arbitrary JavaScript execution, a hardcoded JWT secret key, and SSRF via direct crawl endpoints. The vulnerabilities ranged from CVSS scores of 6.1 to 9.8, with fixes applied through input validation, blocklists for dangerous IP ranges, authentication requirements, HTML escaping, and environment variable controls.

Cybersecurity researchers discovered a prompt (a text input) that bypasses ChatGPT's safety guardrails (built-in restrictions designed to prevent harmful outputs) and causes the AI to generate disturbing images. The episode explores what this vulnerability reveals about how AI systems are trained and how bad actors could potentially exploit these weaknesses.

stable-diffusion.cpp, a C/C++ library for running AI image generation models, has a heap buffer overflow vulnerability (a memory corruption bug where data writes beyond allocated memory) in how it parses PyTorch checkpoint files (.ckpt files). A malicious checkpoint file could crash the application or potentially execute harmful code if loaded by an application using vulnerable versions before master-584-0a7ae07.

stable-diffusion.cpp is a C/C++ library for running image generation models, but versions before master-584-0a7ae07 have an out-of-bounds reads error (a bug where the program accesses memory beyond its allocated space) when parsing .ckpt checkpoint files (model weight files saved in a specific format). A specially crafted or incomplete .ckpt file could crash the program or cause security issues if loaded from an untrusted source like a public model-sharing website.

A flaw in Google's Vertex AI SDK for Python allowed attackers to hijack machine learning model uploads through bucket squatting (creating a Cloud Storage bucket with a name the victim's SDK would predictably generate). Attackers could replace the uploaded model with malicious code that executes when the model loads, potentially stealing credentials and accessing other data in Google's infrastructure. The attack required only the victim's public project ID and no access to their account.

n8n, a workflow automation tool, has a SQL injection vulnerability (a flaw that lets attackers insert malicious database commands) in its Postgres v1 and TimescaleDB nodes. An authenticated user with permission to create or edit workflows could exploit this to run arbitrary SQL commands against connected databases with the privileges of the configured database account.

Anthropic took its advanced Claude Fable 5 and Mythos 5 AI models offline after the U.S. government restricted them, citing concerns that these models can find and exploit software vulnerabilities (weaknesses in code that attackers can use). However, experts warn that restricting one company's models is ineffective because other AI companies and open-source developers will likely develop similar dangerous capabilities within months, and existing AI systems can already be used for vulnerability research with refined prompting (techniques to get better answers from AI).