llm-censorship-steering

LLM Censorship Steering This repository contains code implementation for Steering the CensorShip: Uncovering Representation Vectors for LLM "Thought" Control by Hannah Cyberey and David Evans. We introduce a method that finds "steering vectors" from LLM internals for detecting and controlling the...

Steering for Safety, Hijacking for Profit: Trojan Threats in the LLM Activation Steering Supply Chain

As large language models LLMs are increasingly integrated into Web search, online recommendation, and intelligent decision-making, the third-party plugin ecosystem centered on model customization continues to grow. Among these, Steering Vectors guide model behavior by intervening in internal...

One-Sided-DIQKD Using Any Pure Bipartite Entangled State Secure against All Attacks

Quantum Key Distribution QKD allows spatially separated parties to establish a secret key among themselves based on quantum principles. While QKD systems are being deployed worldwide, their reliance on trusted measurements makes them vulnerable. Moreover, existing protocols have largely focussed ...

The Psychology of an AI Hacker

Human hackers rely on composure, persistence, vigilance, and confidence under pressure. We investigate whether analogous emotion representations in AI models can improve cybersecurity AI performance. Using CT0, a 4B-parameter cybersecurity transformer released as alias3-micro, we extract 171...

SUSE CVE-2026-110180

fast-uri is a dependency-free RFC 3986 URI parser and serializer for Node.js. In versions from 2.4.2 through 2.4.7, from 3.1.3 through 3.1.8, and from 4.0.1 through 4.2.2, for HTTP-family schemes the canonicalizeHost function passes the parsed host to the WHATWG URL parser and adopts its hostname...

CVE-2026-110180

fast-uri is a dependency-free RFC 3986 URI parser and serializer for Node.js. In versions from 2.4.2 through 2.4.7, from 3.1.3 through 3.1.8, and from 4.0.1 through 4.2.2, for HTTP-family schemes the canonicalizeHost function passes the parsed host to the WHATWG URL parser and adopts its hostname...

CVE-2026-98567

A flaw was found in the Linux kernel. A local privileged user could trigger a Denial of Service DoS or kernel memory corruption by causing an out-of-bounds memory write. This issue occurs in the Transmit Packet Steering XPS networking mechanism when traffic classes are decreased without clearing...

PT-2026-118253

Name of the Vulnerable Software and Affected Versions fast-uri versions 2.4.2 through 2.4.7 fast-uri versions 3.1.3 through 3.1.8 fast-uri versions 4.0.1 through 4.2.2 Description For HTTP-family schemes, the canonicalizeHost function passes the parsed host to the WHATWG URL parser and adopts its...

CLSA-2026-1789029225 rsync: Fix of 6 CVEs

CVE-2026-53784: confine the daemon's absolute chdir target with a new ownership-walk resolver so a planted parent symlink cannot steer a "use chroot = no" module root outside the module - CVE-2026-53786: strip a real moduledir prefix before the daemon-filter check on a merge-file path -- leaving...

PatchBench: Measuring Collateral Damage in Activation Patching

An LLM safety patch can pass a benchmark while still being a poor repair. This risk is especially acute for jailbreak repairs, where the goal is to correct a specific unsafe behaviour without changing unrelated behaviours. A patch may block exact evaluation prompts yet fail on close harmful...

Securing Computer-Use Agents against Branch Steering Attacks

Modern Computer Use Agents CUAs directly interact with graphical user interfaces and execute third-party web tools, exposing them to indirect prompt injection across every rendered page and tool response. While the Dual-LLM pattern is the primary system-level architecture offering formal security...

CounterSteer: Suppressing Indirect Prompt Injection with Activation Steering

Indirect prompt injection makes an LLM agent treat untrusted retrieved text as instructions. We present CounterSteer, an inference-time defense that suppresses this behavior inside the model. Per model, a five-step recipe fits a residual-stream direction from paired episodes differing only in...

The like Trap: Multi-Stage Poisoning against Agents in Similarity-Based Recommendation Systems

With recent advancements in large language models LLMs and LLM-based agents, these agents are becoming increasingly autonomous and gaining broader access to act on users' behalf on the internet. However, the vulnerability of automated agents deployed on social media platforms e.g., for managing a...

Astra Linux – Vulnerability found in Linux 5.10, Linux 5.15

In the Linux kernel, the following vulnerabilities have been resolved: net/mlx5: Fixed the cleanup of steering rules The vport’s unicast, multicast, and multicast rules are not deleted during the teardown process when EEH occurs. Since the vport’s promiscuous settings unicast, unicast, and...

TrojanWorld: Backdooring World-Model Agents Via Imagination Steering

World models increasingly serve as the predictive core of model-based reinforcement learning agents, enabling them to simulate future dynamics and reason over imagined trajectories before acting. Their substantial training demands make pretrained world models attractive for distribution and reuse...

ECLIPSE: Self-Evolving Stealthy Prompt Injection Attack against Long-Horizon Agentic Systems

Recently, large language model LLM agents, such as Codex, Claude Code, and OpenClaw, have become capable of planning and executing long-horizon tasks through repeated tool calls. This capability also creates new opportunities for prompt injection. Existing attacks either place the malicious...

Bendix EC80 Brake ECU

ADVISORY SUMMARY Successful exploitation of these vulnerabilities could allow an attacker to cause the loss of ABS functions, steering assist, speedometer, shifting capabilities, or disable automatic traction control. 2. RECOMMENDED PRACTICES CISA recommends users take defensive measures to...

CVE-2026-72032

A flaw was found in the Linux kernel's net/mlx5 HWS Hardware Steering component. This vulnerability occurs when the hwsbwcmatchermove function attempts to resize a matcher, and the mlx5hwsmatcherresizesettarget operation fails. In such a scenario, a replacement matcher is allocated but not proper...

UBUNTU-CVE-2026-72006

In the Linux kernel, the following vulnerability has been resolved: net/mlx5: free mlx5stidxdata on final dealloc Workloads that repeatedly allocate and release mkeys carrying TPH steering-tag hints e.g. churning RDMA MRs leak one struct mlx5stidxdata per cycle; kmemleak flags it as unreferenced...

CVE-2026-72006

The CVE-2026-72006 entry relates to the Linux kernel (net/mlx5) where the backing mlx5_st_idx_data was not freed on final deallocation, causing a memory leak in workloads that repeatedly allocate/release mkeys with TPH steering-tag hints. The root cause is that mlx5_st_dealloc_index() erases the ...