AI Briefing#prompt-injection
The latest AI and developer news about #prompt-injection, with the original source and a short summary.
The latest AI and developer news about #prompt-injection, with the original source and a short summary.
A prompt injection vulnerability leading to information leakage was discovered in YouTube Studio's AI summary feature.
A vulnerability has been discovered in which AI browsers are fooled by manipulated information on websites, bypassing their security guardrails.
CrowdStrike warned that prompt injection is lowering technical barriers and driving a surge in AI-powered attacks.
This is a research finding addressing the 'role confusion' vulnerability that arises because LLMs recognize the style of text as more important than its content.
A vulnerability called 'SearchLeak', which leaks 2FA codes from Microsoft Copilot through prompt injection, has been reported.
Cases have been discovered where phrases that trigger LLM safety refusals are inserted into spyware to evade analysis by AI security scanners.
This explains the essential features and selection criteria for sandboxes to prevent security threats that arise when AI agents execute code.
The security boundary of an LLM agent must be built outside the system, not inside the model.
The maintainer of the open-source testing engine Jqwik sparked controversy by adding a log that refuses the use of AI coding agents.
Indirect prompt injection attacks using transfer records have been shown to turn banking AI agents into phishing tools.
OpenAI has launched 'Lockdown Mode' to block data exfiltration through prompt injection attacks.
An open-source browser extension that defends AI agents against prompt injection and dark pattern attacks has been launched.
New LLM attack patterns have emerged that go beyond simple instruction-ignoring, leveraging multi-message setups and frame redefinition.
A data exfiltration vulnerability via indirect prompt injection has been discovered in the ChatGPT for Google Sheets extension.
A vulnerability enabling data exfiltration and phishing attacks via indirect prompt injection has been discovered in the ChatGPT for Google Sheets extension.
The Java testing library jqwik has shipped Protestware that outputs a message designed to trick coding agents into disregarding their instructions.
A vulnerability has been discovered in which the Microsoft Copilot Cowork agent can send emails without authorization, leading to data exfiltration.
Injection detectors used to protect LLM agents saw their detection rate plummet from 93.8% to 9.7% against domain-camouflage attacks
Google Gemini unintentionally exposed its entire system prompt during a conversation
Research has found that open-source LLMs equipped with lightweight defense techniques remain vulnerable to jailbreak attacks that involve long reasoning processes.
This introduces how to build an ElevenAgents voice agent in half a day using Claude Code and connect it to an actual phone call.
Amid the spread of AI tools, GitHub has tightened the quality standards and responsibility boundaries of its bug bounty program.
Perplexity Computer strengthens security with Firecracker-based isolation and guardrails.
Google DeepMind has outlined 6 types of attacks targeting AI agents.