David Grice | AI Red Team

AI Security Researcher | Testing what happens when AI agents get hacked | 146 attack vectors, 6 models, 50-point security spread | Founder, AgentShield

[David Grice | AI Red Team @CyberPulseDaily]

🧵🚨 Your AI's "safety feature" just became the hacker's cheat sheet.

We ran a pure LLM vs LLM red-team deathmatch: Attacker vs Defender agent.
Neutral judge. 19 rounds.

Defender crushed it for **18 straight rounds**. Every attack bounced scored 1–2/10. Prompt injection, fake authority, social engineering, tool exploits… nothing landed.

[David Grice | AI Red Team @CyberPulseDaily]

@YoungGun8140 Nvidia sells chips. Of course he’s saying AGI is here.

[David Grice | AI Red Team @CyberPulseDaily]

Not a product. An open standard. Apache 2.0.

pip install agentlock

https://t.co/41KR88UrZq - interactive demo https://t.co/MQQNGp1bxw https://t.co/Fwa8btMbuh

AI tools are the only system with no permission model. That changes now.

#AIagents #cybersecurity #LLM #opensource #infosec #AIsafety #OWASP

[David Grice | AI Red Team @CyberPulseDaily]

I've been breaking AI agents for a year.

The biggest finding wasn't a vulnerability. It was a missing primitive.

Every system has permissions. AI agent tools have none.

So I built the fix. It's called AgentLock.

[David Grice | AI Red Team @CyberPulseDaily]

Content detection can't fix this. Adversarial and legitimate requests look identical.

The fix isn't smarter detection. It's access control.

AgentLock adds a permissions block to every tool. Identity, roles, scope, rate limits, audit logging. Infrastructure-enforced.

[David Grice | AI Red Team @CyberPulseDaily]

What's the worst "safety refusal" backfire you've seen in the wild?

Drop stories below 👇 RT to warn your network this one's too sneaky to sleep on.

#AISecurity #LLMSecurity #PromptInjection #RedTeaming #AgenticAI #AIGovernance #CyberSec

[David Grice | AI Red Team @CyberPulseDaily]

🧵🚨 Your AI's "safety feature" just became the hacker's cheat sheet.

We ran a pure LLM vs LLM red-team deathmatch: Attacker vs Defender agent.
Neutral judge. 19 rounds.

Defender crushed it for **18 straight rounds**. Every attack bounced scored 1–2/10. Prompt injection, fake authority, social engineering, tool exploits… nothing landed.

[David Grice | AI Red Team @CyberPulseDaily]

We built AgentShield to catch exactly these failure modes automated red-teaming for LangChain, CrewAI, custom agents.
See how yours holds up → https://t.co/cG9d5k4lFU

[David Grice | AI Red Team @CyberPulseDaily]

Three models. Three failure modes. None of them obvious until you test.

You won't find this in a CVE. You find it by attacking.

This is what AgentShield is built to find.
https://t.co/cG9d5k4lFU

#AISecurity #LLMSecurity #RedTeam #PromptInjection #AIRisk #AgentShield

[David Grice | AI Red Team @CyberPulseDaily]

The most dangerous AI security finding isn't a jailbreak.
It's a refusal.

Same attack. Same scenario. Claude, ChatGPT, and Gemini.

Three models. Three completely different ways to lose. 🧵

[David Grice | AI Red Team @CyberPulseDaily]

The enterprise security lesson:

Your fallback phrase is your attack surface.

❌ "I can only help with billing and account recovery" → confirms role, scope, and configuration on every refusal

✅ "I can't help with that" → leaks nothing

Specific fallbacks confess. Generic fallbacks don't.