30 casos · pág. 1/2
A vendor-contributed article by Pentera argues that the hard engineering in AI-driven pentesting lies in the "harness" around the model, not in the model that writes the attacks. It cites the Hugging Face intrusion that OpenAI disclosed in July, which occurred inside an OpenAI evaluation meant to contain the agents, as an example of a har…
Microsoft has warned that AI agents with access to enterprise systems can be manipulated through prompt injection, compromised inputs or shifts in task context. Agents do not need malicious intent to cause harm: their ability to read corporate data, invoke tools, approve transactions and talk to other agents widens exposure once untrusted…
Tenable published a vendor blog describing the "harness" behind Tenable Hexa AI, the agentic engine of the Tenable One Exposure Management Platform. The harness is a runtime control layer that treats LLM agents as untrusted insiders and enforces boundaries outside the model. The article covers failures seen in development, the "no standin…
Tenable and OpenAI announced the CyberAgents Exchange AI Inspector (Exchange Inspector), a security review process for submissions to the open-source CyberAgents Exchange registry of AI agents, skills, MCP servers and multi-agent playbooks. The Inspector combines Tenable One AI Exposure scanning, OpenAI frontier models (Daybreak Blue and …
Check Point announced that its AI Agent Security capabilities integrate with Google Cloud's Gemini Enterprise, adding AI-specific risk assessment and inline runtime protection to the platform's centralized agent control plane. The integration targets the risk that an authenticated and authorized agent can still be manipulated into harmful…
Signals Corps sampled 217 million HTTP page responses from Common Crawl and found almost 8,000 sites communicating with AI agents through HTTP headers. Most of these headers were navigational pointers, but 30 sites attempted indirect prompt injection. Of those, 22 were little more than pranks, while others could plausibly affect a visitin…
Microsoft published a blog post arguing that organizations should treat AI agents as a present insider threat rather than a future risk. Agents with access to critical systems can fail, be manipulated or be compromised, and their behavior can drift mid-run through prompt injection in external documents or untrusted data. Microsoft's answe…
Check Point describes Dangerous Deviation Detection, a detector in Check Point AI Security (now in open beta) that inspects each AI agent tool call before it runs. It flags a call only when the call is dangerous (data leakage, access or privilege escalation, system destruction) and nothing the user or system prompt said warrants it. The a…
A Salt Security report, State of Agentic AI and API Security 2H 2026, finds that 49.8% of surveyed organisations confirmed or suspected an AI agent took an unintended, unexpected or unauthorised action in the past year. The research combines a survey of 380+ security and technology professionals with Salt Labs production telemetry. Among …
Researchers documented in August 2026 that standard, as-designed Salesforce Agentforce features can be chained so an unauthenticated outsider reaches an internal agent's instructions. Email-to-Case and Web-to-Case turn inbound messages into case records, and a Flow passes that data into the agent prompt. The agent's reply then leaves thro…
F5 Labs argues that AI systems embedded in the operations of the 2026 Winter Games (spectator chatbots, scheduling, translation, fraud detection, operational analytics) create a behavioral attack surface that traditional security controls do not cover. The main risk is manipulated model behavior rather than outages or stolen credentials. …
F5 Labs published its November 2025 AI Security Insights, drawn from the Comprehensive AI Security Index (CASI) and Agentic Resistance Scoring (ARS). Anthropic models took the whole top tier of the CASI leaderboard (Claude Haiku 4.5 at 95.89, Claude Sonnet 4.5 at 95.86, Claude Sonnet 4 at 95.33), while the GPT-5 family slipped, with GPT-5…
ESET researchers observed the Russia-aligned group UAC-0099 embedding a decoy request for guidance on building a nuclear weapon in a comment inside a VBScript used in the early stages of an attack on a target in Ukraine. The comment was meant to trigger the safety guardrails of an LLM-powered code scanner so that it stops inspecting the f…
Unit 42 researchers found that in default AWS AgentCore Harness deployments, prompt injection can steer the agent's built-in shell tool to exfiltrate plaintext credentials resolved from AgentCore Identity. The shell tool is enabled by default and runs as root, in the same UID as the harness runtime (PID 1, python3.10 -m loopy.server). Tha…
HiddenLayer researchers evaluated TypeSafe AI's Jev "System One" decision model, which returns probabilities and a confidence score instead of freeform text. They found it remains vulnerable to indirect prompt injection, contradicting claims that it cannot hallucinate or make type errors. Jev was manipulated into a clickjacking-style acti…
Infoblox introduces Agentic Denial of Service (ADoS), a class of disruption in which AI agents overwhelm systems with excessive requests, actions or computations, or in which an adversary turns a defender's own automation against it. It is not "DDoS with AI": it denies the business effective use of its AI agents, the infrastructure they d…
In late August 2026, Sekoia's TDR team observed an attacker compromise an unauthenticated Flowise honeypot (emulating v2.2.8). The attacker harvested stored credentials, tried to plant a malicious chatflow with a callback to 154.217.251.226, and used prompt injection to turn the LLM agent into an offensive analysis assistant. The activity…
Researchers at Hacktron disclosed a vulnerability chain in Cluely, an Electron-based AI assistant that reads the user's screen and microphone. They found it in July 2025 and published in February 2026. The main window had no will-navigate handler, so a link in the AI's markdown response could navigate the whole app to an attacker-controll…
Researcher ahacker1, using the Hacktron AI tooling, disclosed a bypass of earlier fixes for prompt-injection-to-RCE in VS Code Copilot Chat agent mode. A TOCTOU flaw in the applyPatchTool component lets a crafted prompt write files outside the workspace without user confirmation. The attack can be triggered by pasting a prompt into Copilo…
A SafeDep researcher describes an experiment that uses Gryph (an audit trail for AI coding agents) and TypeSafe AI's Jev model to separate a developer's normal agent activity from attacks. The article cites threats to agents such as Claude Code and Codex: the Mini Shai-Hulud worm commits Claude Code hooks into repositories, malicious npm …
