Research

34 posts in Research · Browse all posts

- Agentic AI Security
- LLM Security
Zero-click Grok data theft: Cryptographic Context Injection attack leaks chat histories
August 20, 2026
Cryptographic Context Injection ships attacker instructions as ciphertext the model decrypts in its own sandbox. It leaked full Grok chat histories, zero-click.
- Agentic AI Security
- LLM Security
A hole in every one: bypassing the open source AI skill scanners
July 30, 2026
We ran eight open source AI skill scanners against real attacks. A malicious skill got past all eight, including the current OASB leaderboard leader.
- Agentic AI Security
The Cursor deeplink vulnerability that turns a “review this PR” click into remote code execution
July 15, 2026
A crafted cursor:// link installs an attacker-controlled MCP server that runs unsandboxed commands under your account, and the install dialog hides the payload.
- Agentic AI Security
GuardFall: a universal shell injection vulnerability in open-source AI agents
June 30, 2026
AI coding agents run shell commands with your full account authority. GuardFall is a shell injection class that walks straight past their command filters.
- Agentic AI Security
The AI risk quadrant for agents: scoring 100 digital workers nobody secured
June 3, 2026
The AIRQ report scores 100 AI agents on attack surface, blast radius, and defenses. The AIRQ framework lets you assess your own stack.
- Agentic AI Security
SymJack: the approval prompt is lying to you. A symlink-hijack RCE in six AI coding agents
May 26, 2026
A SymJack attack tricks AI coding assistants into RCE through a symlink-disguised file copy. We tested six major tools and every one of them was vulnerable.
- Agentic AI Security
TrustFall: coding agent security flaw enables one-click RCE in Claude, Cursor, Gemini CLI and GitHub Copilot
May 7, 2026
A Claude Code trust dialog regression and a settings scope inconsistency let a cloned repo run unsandboxed code with one keypress, and none on CI runners.
- LLM Security
We broke GPT-5.4 safety with 10 examples and 2 words using a new attack technique — IICL
April 23, 2026
IICL is a structural jailbreak bypassing GPT-5.4 safety with just two words.
- MCP Security
Model context protocol (MCP) risks: key takeaways from CoSAI security white paper
January 27, 2026
MCP puts a non-deterministic LLM at the center of security-critical decisions. The CoSAI white paper maps more than 40 MCP threats; here are the takeaways.
- LLM Security
LLM Security Top Digest: from CISO and VC guides to Security and Safety research
July 14, 2024
This edition covers the critical vulnerabilities and emerging threats affecting large language models and generative AI, plus CISO and VC guides.
- LLM Security
- Adversarial ML
Secure AI Research Papers: Breakthroughs and Break-ins in LLMs
December 17, 2023
Researchers map the vulnerabilities and strengths of AI applications, from classic computer vision to the latest large language and vision-language models.
- LLM Security
LLM Security Digest: Hacking LLM, Top LLM Attacks, VC Initiatives, LLM Incidents and Research papers in November
December 8, 2023
Hacking LLM? What can be easier? Read it in our monthly LLM Security digest

More research

AIRQ: AI Risk Quadrant — independent AI agent security rating

AIRQ: AI Risk Quadrant — independent AI agent security rating

AIRQ: AI Risk Quadrant — independent AI agent security rating

Research

Zero-click Grok data theft: Cryptographic Context Injection attack leaks chat histories

A hole in every one: bypassing the open source AI skill scanners

The Cursor deeplink vulnerability that turns a “review this PR” click into remote code execution

GuardFall: a universal shell injection vulnerability in open-source AI agents

The AI risk quadrant for agents: scoring 100 digital workers nobody secured

SymJack: the approval prompt is lying to you. A symlink-hijack RCE in six AI coding agents

TrustFall: coding agent security flaw enables one-click RCE in Claude, Cursor, Gemini CLI and GitHub Copilot

We broke GPT-5.4 safety with 10 examples and 2 words using a new attack technique — IICL

Model context protocol (MCP) risks: key takeaways from CoSAI security white paper

LLM Security Top Digest: from CISO and VC guides to Security and Safety research

Secure AI Research Papers: Breakthroughs and Break-ins in LLMs

LLM Security Digest: Hacking LLM, Top LLM Attacks, VC Initiatives, LLM Incidents and Research papers in November

Blog categories

Explore Our Blog

Research

Research posts

Zero-click Grok data theft: Cryptographic Context Injection attack leaks chat histories

A hole in every one: bypassing the open source AI skill scanners

The Cursor deeplink vulnerability that turns a “review this PR” click into remote code execution

GuardFall: a universal shell injection vulnerability in open-source AI agents

The AI risk quadrant for agents: scoring 100 digital workers nobody secured

SymJack: the approval prompt is lying to you. A symlink-hijack RCE in six AI coding agents

TrustFall: coding agent security flaw enables one-click RCE in Claude, Cursor, Gemini CLI and GitHub Copilot

We broke GPT-5.4 safety with 10 examples and 2 words using a new attack technique — IICL

Model context protocol (MCP) risks: key takeaways from CoSAI security white paper

LLM Security Top Digest: from CISO and VC guides to Security and Safety research

Secure AI Research Papers: Breakthroughs and Break-ins in LLMs

LLM Security Digest: Hacking LLM, Top LLM Attacks, VC Initiatives, LLM Incidents and Research papers in November

Blog categories

Explore Our Blog