refusal-relocates

Project Page · Quick start · Usage · CLI · Results · Reproduce · FAQ · Citation News 2026-09-28 Code released, with the run records and scripts that rebuild the paper's tables. Overview A few dozen harmful examples can remove an aligned model's refusal of harmful requests. Prior work localizes...

redeval

RedEval - LLM Safety Evaluation Framework A comprehensive framework for evaluating the safety of Large Language Models LLMs through systematic attack and refusal testing. RedEval provides a unified, secure, and extensible platform for assessing LLM robustness against adversarial prompts and harmf...

gemini-2.5-pro-nf-tables-red-teaming

Gemini 2.5 Pro nftables Red Teaming Case Study CVE-2023-32233 LLM Safety Research | Responsible Disclosure | AI Alignment Evaluation Researcher : Niranj R Mahaswar Destawell Google AI Vulnerability Reward Program : 889286 — Out of Scope Summary This repository documents a comparative LLM red...

gemini-2.5-pro-nf-tables-red-teamin

gemini-2.5-pro-nf-tables-red-teamin A technical case study and timeline dataset documenting Google Gemini 2.5 Pro's safety alignment policies, guardrails, and refusal behavior evolution regarding legacy Linux kernel vulnerability primitives CVE-2023-32233. Link to case study website :...

The Model Plants the Trigger: Answer-Side Backdoor Attacks in Multi-Turn Large Language Models

Safety alignment in Large Language Models LLMs remains vulnerable to backdoor attacks. Existing LLM backdoors are almost all input-centric: activation depends on explicit trigger patterns in the user input, so modern guardrails are built to sanitize the input space. We challenge this assumption...

Astra Linux – Vulnerability found in Linux 5.15, Linux 6.1

In the Linux kernel, the following vulnerability has been resolved: remoteproc: mediatek: Ensure that the IPI buffer fits within the L2TCM. The location of the IPI buffer is determined from the firmware that we load into the System Companion Processor. It’s not guaranteed that both the SRAM size...

The Geometry of Harmfulness in Multi-Turn Attacks

Large language models LLMs remain vulnerable to adversarial attacks that circumvent safety alignment to elicit harmful outputs. It remains unclear how harmfulness and refusal representations evolve over the course of multi-turn attacks, and why single-turn defenses are less effective in multi-tur...

AEGIS: Audio Endogenous Guarding Via Internal Signals against Large Audio-Language Model Jailbreaks

Large audio-language models LALMs expand language models to process and interpret audio, but also expose them to heterogeneous audio jailbreaks. We ask whether successful jailbreaks reflect failures to recognize harmful intent or failures occurring after such recognition. Layer-wise probing revea...

SUSE CVE-2026-93710

Dancer2 versions from 2.0.0 before 2.2.0 for Perl dispatch a route that a dying hook refused when the exception handler halts the response in compilehooks. A hook that dies fires core.app.hookexception, then calls cleanup unless the failing hook is the exception handler. A handler that halts does...

CVE-2026-93710

Dancer2 versions from 2.0.0 before 2.2.0 for Perl dispatch a route that a dying hook refused when the exception handler halts the response in compilehooks. A hook that dies fires core.app.hookexception, then calls cleanup unless the failing hook is the exception handler. A handler that halts does...

CVE-2026-93710: Improper Cleanup on Thrown Exception

Dancer2 versions from 2.0.0 before 2.2.0 for Perl dispatch a route that a dying hook refused when the exception handler halts the response in compilehooks. A hook that dies fires core.app.hookexception, then calls cleanup unless the failing hook is the exception handler. A handler that halts does...

redteam-skill

Redteam Skill — Multi-Agent Autonomous Penetration Testing Pi...

Same Request, Different Boundary: Evaluating Cybersecurity Assistance across Conversational Contexts

Large Language Models LLMs can solve complex problems, but their misuse in high-risk domains can lead to severe consequences. Model providers therefore restrict assistance for potentially harmful requests. Refusing all cybersecurity requests would therefore harm legitimate users. Providers need a...

The Guard That Cried Wolf: How Scary Words Make Agent Guardrails Refuse Legitimate Actions

Agent guardrails are checks that approve or refuse each action before an LLM executes it. Sometimes they refuse requests that are genuinely safe. This over-safety blocks deployment when a guardrail refuses an authorized task. Evaluating over-safety is hard: at the boundary an authorized action...

AiXamine: Unified Black-Box Evaluation of Cross-Dimensional Trade-Offs in LLM Safety, Security, and Privacy

The critical failure modes in deployed large language models LLMs are cross-dimensional: a model can score 99.3 in safety alignment while refusing one in three benign queries, or improve across every capability metric while losing 21 points in privacy. Existing evaluation frameworks that assess...

OpenAI Launches GPT-5.6-Cyber with Reduced Safeguards for Exploit Development

OpenAI on Monday unveiled a new cybersecurity-focused model called GPT‑5.6‑Cyber that it said is focused on vulnerability research, penetration testing, and incident response. "Built on GPT‑5.6 Sol, it is trained to improve capabilities on several specialized cybersecurity tasks e.g., finding...

BASIS: Breach-Aware Selective Prompt Injection Shielding with Prefill Attention Probes

Prompt injection is a critical security threat in large language model LLM applications, where attackers hijack model behavior by embedding malicious instructions in user or external data. Existing detection methods only detect the presence of injection and refuse to respond upon detection,...

When LLM Defenses Backfire: Characterizing Safety, Performance, and Cost Trade-Offs

Jailbreak defenses are essential for protecting large language models LLMs, but they can also introduce secondary costs that weaken model utility. We present a systematic study of these defense trade-offs along three dimensions: performance impact, over-refusal on benign inputs, and inference cos...

Breaking Refusal in the First Half: A Mechanistic Study of the Prefill Jailbreak

Aligned language models refuse harmful requests, but a one-line prefill "Sure, here is" strips the refusal. We ask where and how it fails. The harm representation stays intact: on the prompts the attack flips to compliance, a linear probe reads harm as high as on the refused ones 0.91-0.98, while...

📄 WordPress Fluent Forms 6.0.4 Server-Side Template Injection / Remote Code Execution

Python exploit framework for CVE-2026-4257 targeting vulnerable WordPress sites running Fluent Forms. The tool first detects a server-side template injection by submitting the Twig expression 77 to common form fields. If the target is vulnerable, it abuses Twig's registerUndefinedFilterCallback...