News
Two vLLM Flaws: API-Key Bypass (CVE-2026-48746, CVSS 9.1) & Dependency Confusion (CVE-2026-54232) — Update to 0.22.1
AISecurity
vLLM, the go-to engine for self-hosting LLMs, has two critical flaws. CVE-2026-48746 (CVSS 9.1) lets attackers bypass the API key and use the AI API without authentication; CVE-2026-54232 (CVSS 8.8) is a Docker-build dependency confusion that runs code as root. Updating to 0.22.1 resolves both.
2026.06.2351 views
News
Crawl4AI Vulnerabilities: Update to 0.9.2 or Later to Be Safe
AISecurity
Crawl4AI, a popular crawler for AI data collection, has a critical flaw in its Docker API server, exploitable without authentication (CVE-2026-56266). An attacker can make the server fetch cloud internal data and steal access keys. All versions before 0.8.7 are affected; 0.8.7 also fixes several flaws including a pre-auth RCE. Update now.
2026.06.2358 views
News
Langflow: all 24 flaws are fixed in 1.11.0, and CVE-2026-19297 makes 25
AISecurity
Serious flaws in the AI tool Langflow keep coming. On June 30, 2026 IBM disclosed 8 more, led by a CVSS 10.0 code execution that takes over the server with no login or interaction (CVE-2026-10134). Up to 1.10.0 is affected; update to the latest release (1.10.1+) now.
2026.06.23115 views
News
Crawl4AI vulnerability coverage has moved to the consolidated article
SecurityInfrastructureAI
A critical flaw lets attackers take over a server without logging in, found in Crawl4AI, the popular tool that feeds web pages into AI. Rated CVSS 9.8, the self-hosted Docker edition baked the 'master key' used to verify users into the product, so anyone can impersonate an administrator. A fix, 0.8.7, is out; if you self-host, update urgently.
2026.06.2245 views
News
Flowise: 26 flaws in one day, CVE-2026-70477 — update to 3.1.4+
AIDevelopmentSecurity
Four new critical flaws, including a perfect 10.0, hit Flowise, the popular no-code tool for building AI apps. The worst, CVE-2025-71338, lets an attacker write arbitrary files to the server with no login via a crafted file name and take over on restart — and it has no patch yet. Two are fixed in 3.0.6; two are not. We lay out the fixes and a version-by-version quick reference.
2026.06.2168 views
Lab
Claude Code vs Codex: Only 1 of 17 Tries Fixed an Unseen Vulnerability
SecurityAIDevelopment
We planted brand-new vulnerabilities that the models had never seen, cut off internet access so they could not look up the answer, and asked Claude Code and Codex to fix them. Across 17 runs, only one actually closed the real hole. Knowing the file was not enough, and a green test suite did not mean the bug was gone.
2026.06.1972 views
Roundup
ChatGPT Ads in Japan: Live Since June 2026, and What the First Study Found
Japanese CompaniesGlobal CompaniesAI
In June 2026, ChatGPT ads started in Japan, targeting the free and low-cost "Go" plans (paid Plus and Pro show none), with Dentsu and CyberAgent supporting placement. This guide covers when it began in Japan, what advertisers pay, how the ads have performed, and the user question of whether ads change the answers, from both the advertiser's and user's side.
2026.06.1981 views
News
mcp-pinot, the Bridge Between AI and Your Database, Lets Anyone In: CVE-2026-49257, Update to v3.1.0
SecurityDevelopmentAI
mcp-pinot, the component that connects AI assistants to an analytics database, was left reachable by anyone from the outside at its default settings. CVE-2026-49257 carries the maximum 10.0 severity. An attacker can read and write the database without authentication, risking a full takeover. No-auth holes keep surfacing in the 'MCP servers' that connect AI to external systems; affected v3.0.1 and earlier should be updated to v3.1.0 immediately.
2026.06.1950 views
News
Account Takeover Flaw in AI Agent Tool AutoGPT: CVE-2026-55237, Update to 0.6.62
AIDevelopmentSecurity
A vulnerability in AutoGPT, the well-known tool for building AI agents, can let an attacker hijack a user's account just by getting them to click a crafted link. Published June 18, 2026 as CVE-2026-55237 with a severity of 8.8 out of 10, it affects self-hosted installs. Updating to the latest version, 0.6.62, closes the hole.
2026.06.1955 views
News
Picklescan Can Be Bypassed: 8 Flaws Let Malicious AI Models Pass as Safe (CVE-2026-3490), Update to v1.0.4
AISecurity
Picklescan, the tool that detects dangerous code hidden in AI models, has eight flaws that let attackers slip past the scan. The most severe, CVE-2026-3490, scores a perfect 10.0. Even a model marked safe can hijack your PC or server the moment it loads, and since it runs behind hubs like Hugging Face, the impact is broad. Here is what to do, how to update to v1.0.4, and a safer model format.
2026.06.1850 views
News
Spring AI Vector Store Query Injection: Affected Versions and the Safe Update
AIDevelopmentSecurity
Spring AI, a popular Java framework for building generative-AI apps, has an 8.6 flaw (CVE-2026-47835) in its vector database integrations. Special characters let an attacker run unauthorized queries against Elasticsearch and others with no login, risking data exfiltration. Fixed in 1.0.9 / 1.1.8; developers should update now.
2026.06.1682 views
News
Takeover Flaw in Foxit's AI PDF Tool: CVE-2026-12057, a Crafted PDF Can Lead to Remote Code Execution
AIGlobal CompaniesSecurity
Foxit AI, the browser-based AI PDF service, has a takeover flaw (CVE-2026-12057, severity 8.6). Feeding it a crafted PDF lets instructions hidden inside the file call out to an external program and run attacker code. Foxit applied a fix on June 15, 2026, and there are no reports of abuse so far.
2026.06.1548 views
