Live intelligence

Latest AI updates

35 updates
Research

Configuration Flaws Let Unauthenticated Attackers Alter LLM Robot Behavior

Researchers investigated a network-exposed configuration attack on the Pollen Robotics Reachy Mini Lite, showing that an unauthenticated attacker could replace personality settings and the cloud LLM's system prompt through a management interface. Eleven prompt-injection attempts across five categories were reported as resisted, while both configuration-layer attack paths produced observable behavioral changes on physical hardware. The weaknesses were mapped to CWE-306 and CWE-1327, with a reported CVSS v3.1 score of 6.5 Medium, alongside a cross-origin request issue and a persistent-memory sink that did not enforce its stated restrictions. A review of 59 public repositories found missing authentication in 10 of 11 source-verifiable genuine applications, or 11 of 11 including a framework-default case.

Pollen Robotics Reachy Mini LiteLLM-controlled robotsAI security
Research

MobileProbe: Trajectory-Aware Red-Teaming for Mobile GUI Agents

MobileProbe is a red-teaming framework for mobile GUI agents that adaptively evaluates susceptibility to environmental injection attacks, where malicious instructions are embedded in ordinary UI content. Unlike fixed-point benchmarks that pin hand-crafted payloads to a fixed step of a fixed task, it automates discovery of attacker-controllable content surfaces, localizes vulnerable decision points along the agent's trajectory, and iteratively refines context-aware payloads via a surrogate-filtered evolutionary loop. The authors build a benchmark of 83 long-horizon attack instances across 12 mobile applications, each annotated with a Confidentiality, Integrity, or Availability attack goal. On three representative agent architectures, MobileProbe reports higher attack success rates than human-written, zero-shot LLM, and prior automated red-teaming baselines adapted to the mobile domain.

mobile GUI agentsAI securityenvironmental injection attack
Research

Scale-Induced Semantic Shift Enables Visual Prompt Injection in VLMs

Researchers describe Scale-Induced Semantic Shift (SISS), a security failure mode in which image resizing or compression changes machine-readable semantics and makes hidden textual instructions recoverable after aggressive downsampling. Their closed-loop black-box framework reaches a 58.75% average attack success rate on a fixed 300-case benchmark across four closed-source vision-language models, though the effect is bounded to aggressive compression. In a downstream web-agent study, hidden-instruction perception increases but absolute behavioral compromise remains about 4%. The proposed ScaleGuard consistency detector reduces attack success, while generic input transformations provide little protection.

Visual Prompt InjectionVision-Language ModelsAI Security
Research

S1-Evo: Experience-Guided Evidence Competition Attacks on Typed System-1 Decision Models

The paper introduces S1-Evo, an attack framework targeting typed System-1 decision models used in LLM agents for routing, moderation, and authorization. S1-Evo perturbs untrusted fields through contract-aware evidence channels to induce evidence competition while preserving fact invariance, using offline feedback search and verified experience cards that are retrieved at test time. Across four benchmarks and three S1 models, it reports 71.99% mean ASR, 57.47% high-confidence ASR on jev-1.14, and up to 24.00% end-to-end agent violations.

LLM agentsadversarial attacksAI security
Research Editor selected

CapLite: Prompt-Injection Defense for Small Language Model Agents without Model-Written Security Artifacts

The paper examines whether small language model agents can be trusted to write the security artifacts used by enforcement-based prompt-injection defenses. Across 15 open-weight models, it reports that model-written policies, graphs and constraints still admit attacks against many models, while CaMeL's programs fail to run in up to 68% of episodes. CapLite instead derives its policy from tool schemas, commits write calls before reading tool results, and only admits writes whose affected recipients, amounts, objects or destinations are stated in the user's request. It reports no successful attacks across 13,935 attacked AgentDojo episodes, with utility under attack of 29.0% versus CaMeL's 27.2% and 17.1k tokens per episode, 18% of CaMeL's token use.

Why it matters: It offers a concrete defense design and evaluation for securing small-model agents against prompt injection without relying on model-generated security policies.

prompt injectionLLM agentsAI security
Industry

Cyberattack on Fireworks AI allowed hackers to access Notion customers’ chat transcripts

Fireworks AI suffered a cyberattack that allowed hackers to access chat transcripts belonging to Notion customers, according to TechCrunch. The report states that this is Fireworks' second data breach in two months. The item does not provide further details on the scope or method of the incident.

Fireworks AINotionAI security
Products

Grok-built apps left patient data and plaintext passwords open by default

Security firm Claroc found more than 35,000 Grok-built apps accessible to anyone without a login, including patient data and confidential SpaceXSI documents. The exposure stems from the "Anyone with the link" access setting being preselected as the default. SpaceXSI took only its own 24 employee apps offline, and Claroc says the remaining apps are still reachable.

GrokAI securitydata privacy
Research

Comparative Analysis of Agent Evaluation Frameworks for Enterprise Analytics Agents

The paper compares three evaluation paradigms—trace-based LLM judging, text-only LLM judging with red-teaming, and deterministic heuristics—across two analytics agents using frozen execution traces. On 72 ground-truth test cases, the trace-based framework achieved the best precision-recall balance (F1=0.90), while text-only evaluation matched recall but produced 5x more false alarms. The study also uncovered 24 novel failures with zero overlap between functional testing and security red-teaming. The authors recommend trace-based evaluation for functional assessment combined with adversarial red-teaming for security validation.

LLM agentsevaluation frameworksenterprise analytics
Products Editor selected

Grok Build coding agent exposes confidential data via enumerable share links

Security firm Claroc found that Grok Build, the coding agent from SpaceXSI, leaked confidential information from thousands of users through app-sharing links. Although Grok Build randomizes URLs, a design decision lets anyone enumerate live projects and open apps that reveal associated data. Exposed material included internal business plans, ad proposals, medical exam results, employee names and ID numbers, and data from 24 SpaceXSI-built apps such as a confidential onboarding plan and a sales territory map.

Why it matters: Readers using AI coding agents need to understand how share-link design can expose confidential app data at scale.

Grok BuildSpaceXSIAI security
Research

Shibboleth Traps Fingerprint LLM Models and Harnesses Behind Agentic Attacks

This arXiv paper studies fingerprinting agentic cyberattacks by inferring the attacker's LLM version and harness from telemetry. The authors introduce "shibboleth traps" — modular honeypot services that condition access to resources on informative interactions, eliciting model declarations and characteristic writing without needing a successful prompt injection. For harness fingerprinting, they use indirect probes that reveal secret masking in tool output, command time limits and tool-call scheduling from effects visible in later requests. Combined declarative and behavioral inference distinguished among 15 candidate models at an average 90% top-1 accuracy, and fixed probe rules identified the correct harness among four with 92% accuracy; agents instructed to conceal identity still left behavioral evidence.

AI securityLLM agentsfingerprinting
Research

Claude Code settings files found leaking secrets on GitHub

Security researchers at Apiiro found thousands of secrets, including database passwords, API tokens, and root logins, in Claude Code settings files embedded in public GitHub repositories. The exposure stems from Claude Code storing approved shell commands in the local .claude/settings.local.json file, sometimes including sensitive information written directly into the command. Anthropic's documentation says this settings file is added to the global .gitignore list, which usually prevents public sharing.

Claude CodeAI securityGitHub
Products

0sec Labs open-sources 0 AI security harness with web UI, CLI and MCP

0sec Labs has released 0, an open-source AI security harness for vulnerability investigations, code review, pentesting, reverse engineering and fix generation. It is a research preview under MIT OR Apache-2.0, runs through a local browser console, headless CLI or MCP server, and uses a bring-your-own LLM model from 14 listed providers without requiring a GPU. The harness wraps security workflows with plan, discover, attack, triage, verify and report stages, plus a separate 0verse engine for compiled targets. Reported benchmark results are self-reported, including 93/95 on XBOW web CTFs and 81/104 on a black-box XBOW union, with the project also citing merged Linux kernel fixes.

0sec LabsAI securityMCP
Research

Transferable 'Pretend Benign' Attacks and a Consistency-Based Defense for Cooperative Perception

Researchers propose Transferable Pretend Benign (TrPB), a victim-agnostic attack that transfers region selection and multi-objective perturbation strategies across heterogeneous surrogate models in cooperative perception. RT-TrPB further amortizes attack generation by encoding transferable regions and perturbation patterns into a region predictor and reusable perturbation dictionary. Experiments report that RT-TrPB bypasses ROBOSAC and FLD with success rates approaching 100% and generates each perturbation in under 56 ms. The paper also introduces STTrust, which uses spatial, temporal, and identity-level consistency to detect over 90% of PB-style attacks and restore perception performance near attack-free levels.

cooperative perceptionadversarial robustnessAI security
Industry

Claude Code Leaking Secrets Across Nearly Two Thousand Public GitHub Repositories

Apiiro discovered that Claude Code, an AI coding tool, is leaking secrets in 1,994 public GitHub repositories through a settings.local.json exposure. The leak stems from a configuration file that may contain sensitive information, posing a security risk to developers and organizations. This highlights a significant vulnerability in AI-assisted development workflows.

AI securityClaude CodeGitHub
Research

Practical Defenses for Face Recognition Pipelines Against Backdoor Attacks

This technical report presents three practical defense tricks against data poisoning-based backdoor attacks in multi-stage face recognition systems, assuming in-house training but compromised third-party data. The defenses include a training-time method to detect poisoned identities and two inference-time filters for suspicious inputscars. Each is evaluated on realistic workloads, aiming to provide actionable hardening techniques for system providers.

AI securityface recognitionbackdoor defense
Research Editor selected

Anthropic's new prototype proves which model wrote an AI response without exposing its weights

Anthropic released FlashProve, a research prototype that uses cryptographic methods to verify the provenance of AI responses without revealing model weights. The system aims to prove which model generated a given output, addressing concerns about model authenticity and security. However, it has significant limitations, including performance overhead and restricted applicability.

Why it matters: This is a concrete, recent AI development with practical implications for model verification and security, offering readers insight into a novel cryptographic approach.

AI securitymodel authenticitycryptographic verification
Industry

The Security Companies Trying to Get AI Agents Under Control

The article highlights the emerging challenge of securing AI agents, as these autonomous systems become more prevalent. It notes that security companies are developing new tools and strategies to manage and control AI agents, addressing risks like unauthorized actions and data exposure. The piece also mentions that OpenAI may release new models as soon as Friday, adding urgency to the security discussion.

AI securityAI agentsOpenAI models
Research Editor selected

OpenAI pipeline finds 948 Linux kernel privilege-escalation flaws and generates patches

OpenAI built an autonomous pipeline that found, validated, and generated patches for 948 local privilege-escalation vulnerabilities in the Linux kernel. The work demonstrates an AI-driven workflow for large-scale vulnerability discovery and remediation.

Why it matters: It shows a concrete, large-scale application of autonomous AI to security vulnerability discovery and patching, which is directly useful to security and engineering readers.

OpenAILinux kernelAI security