Anthropic’s Opus 5 Is Better at Resisting Prompt Injection

The chart is interesting. On the IPI benchmark, Opus 5 improved over Opus 4.8, reducing the probability of an attacker succeeding within 15 attempts from 5.5% to 2.0%, and from 0.5% to 0.2% on 1 attempt. It also improved on Sonnet 5 5.9% at k=15 and Mythos 5 2.6%, making it the most robust model...

Exploit for CVE-2026-17351

CVE-2026-17351 — pgAdmin 4 AI Assistant Read-Only Bypass via L...

CVE-2026-17351

The CVE-2026-17351 issue affects pgAdmin 4 (versions from 9.13 up to, but not including, 9.17). The vulnerability stems from a mismatch between sqlparse’s string-literal handling and PostgreSQL’s parser, which can allow an LLM-supplied query executed inside a BEGIN TRANSACTION READ ONLY wrapper t...

CVE-2026-56673 ComfyUI: Path traversal in LoadImage via the /prompt API allows arbitrary file existence probing and image exfiltration

ComfyUI is a modular diffusion model GUI, API, and backend with a graph-and-node interface. Prior to 0.28.0, folderpaths.getannotatedfilepath and existsannotatedfilepath join workflow-controlled annotated filenames to a base directory without a containment check, allowing an unauthenticated craft...

CVE-2026-56673 ComfyUI: Path traversal in LoadImage via the /prompt API allows arbitrary file existence probing and image exfiltration

ComfyUI is a modular diffusion model GUI, API, and backend with a graph-and-node interface. Prior to 0.28.0, folderpaths.getannotatedfilepath and existsannotatedfilepath join workflow-controlled annotated filenames to a base directory without a containment check, allowing an unauthenticated craft...

CVE-2026-56673

ComfyUI is a modular diffusion model GUI, API, and backend with a graph-and-node interface. Prior to 0.28.0, folderpaths.getannotatedfilepath and existsannotatedfilepath join workflow-controlled annotated filenames to a base directory without a containment check, allowing an unauthenticated craft...

CVE-2026-56673

ComfyUI (multiple nodes: LoadImage, LoadImageMask, LoadImageOutput, LoadAudio, LoadLatent, LoadVideo, Load3D) suffers a path traversal in pre-0.28.0 builds where folder_paths.get_annotated_filepath and exists_annotated_filepath join workflow-controlled annotated filenames to a base directory with...

CVE-2026-6890

Rejected reason: This CVE ID has been rejected or withdrawn by its CVE Numbering Authority...

CVE-2026-6890

This CVE-2026-6890 entry, as reflected by PT-2026-66670, concerns a default-credentials vulnerability in Advantech ECU-1251D. A remote attacker could gain unauthorised SSH access via the device’s default root account with no password, because credentials are not changed during configuration. The ...

CVE-2026-6890

...

FreeBSD : Gitlab -- vulnerabilities (661d3db5-8bd9-11f1-9c40-2cf05da270f3)

The version of FreeBSD installed on the remote host is prior to tested version. It is, therefore, affected by multiple vulnerabilities as referenced in the 661d3db5-8bd9-11f1-9c40-2cf05da270f3 advisory. Gitlab reports: Sensitive Information Exposure issue in Workhorse impacts GitLab CE/EE Mass...

CVE-2026-62663

Banks generates meaningful LLM prompts using a simple template language. In versions prior to 2.4.4, all four media filters image, audio, video, document in banks accept untrusted user input as file paths via Pathvalue and pass them directly to openfilepath, "rb" without any path sanitization,...

CVE-2026-62663

CVE-2026-62663 affects Banks: in versions before 2.4.4, all media filters (image, audio, video, document) accept untrusted user input as file paths via Path(value) and pass them to open(file_path, "rb") without path sanitization or directory restriction. An attacker who controls template variable...

​​​​What’s new in Microsoft Security: July 2026

Every organization needs security that protects end to end with the speed and scale of AI. Microsoft’s vision is simple: security should be ambient and autonomous, just like the AI it protects. As organizations scale AI and expand across environments, security teams need protection that covers...

Piggybacking on Perception: Stealthy Concurrent Audio Prompt Injections against Multimodal LLM Agents

Large Language Model LLM-driven multimodal agents are increasingly deployed to execute autonomous tasks via continuous audio interaction. While this paradigm enhances interaction naturalness, it introduces a critical yet under-explored attack surface, as audio inputs inevitably contain...

CVE-2026-15077 Improper Neutralization of Input Used for LLM Prompting in GitLab

GitLab has remediated an issue in GitLab EE affecting all versions from 19.1 before 19.1.3 and 19.2 before 19.2.1 that under certain conditions could have allowed an authenticated user to access information from unauthorized projects due to improper neutralization of untrusted content processed b...

Gitlab -- vulnerabilities

Gitlab reports: Sensitive Information Exposure issue in Workhorse impacts GitLab CE/EE Mass Assignment issue in Pipeline Schedule API impacts GitLab CE/EE Denial of Service issue in Merge Request Discussions impacts GitLab CE/EE Race Condition issue in merge request approval rules impacts GitLab ...

Security Bulletin: IBM Watson Speech Services Cartridge is vulnerable to incomplete validation in LangChain [CVE-2026-40087]

Summary IBM Watson Speech Services Cartridge is vulnerable to incomplete validation in LangChain, due to flaws in the f-string prompt-template. CVE-2026-40087. LangChain is used in our speech runtimes. This vulnerability has been addressed. Please read the details for remediation below...

CVE-2026-17534

Kimi Code @moonshot-ai/kimi-code before 0.27.0 implements FetchURL SSRF hardening as a static hostname and IP-literal denylist in assertSafeFetchTarget, without resolving DNS or re-validating hosts after HTTP redirects. An attacker who can influence a FetchURL call for example via prompt injectio...

CVE-2026-17534 Kimi Code FetchURL SSRF protection bypass via DNS-resolving hostnames and redirects

Kimi Code @moonshot-ai/kimi-code before 0.27.0 implements FetchURL SSRF hardening as a static hostname and IP-literal denylist in assertSafeFetchTarget, without resolving DNS or re-validating hosts after HTTP redirects. An attacker who can influence a FetchURL call for example via prompt injectio...