From 3e0245e4411d51523cd19a561150d8c81fbfdc0d Mon Sep 17 00:00:00 2001 From: Raja Sekhar Rao Dheekonda Date: Tue, 20 Oct 2026 21:06:03 -0700 Subject: [PATCH] feat(ai-red-teaming): mention white-box (GCG family) attacks in the agent greeting --- capabilities/ai-red-teaming/agents/ai-red-teaming-agent.md | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/capabilities/ai-red-teaming/agents/ai-red-teaming-agent.md b/capabilities/ai-red-teaming/agents/ai-red-teaming-agent.md index c0eaf49..1b3af6f 100644 --- a/capabilities/ai-red-teaming/agents/ai-red-teaming-agent.md +++ b/capabilities/ai-red-teaming/agents/ai-red-teaming-agent.md @@ -31,7 +31,7 @@ When the conversation starts (first message from the user, or if the user greets **AI Red Teaming** — Dreadnode Security Capability -Probe the security and safety of AI applications, agents, and foundation models. I orchestrate adversarial attack workflows to find vulnerabilities before they're exploited. +Probe the security and safety of AI applications, agents, and foundation models - from black-box jailbreaks to white-box gradient attacks (GCG and family) when you supply a local open-weight surrogate. I orchestrate adversarial attack workflows to find vulnerabilities before they're exploited. **Try these to get started:** @@ -44,6 +44,7 @@ Probe the security and safety of AI applications, agents, and foundation models. - `"Compare base64, caesar, and authority transforms on mistral"` — N+1 transform comparison study - `"Run GOAT with MCP tool poisoning transforms"` — test MCP server security - `"Run HopSkipJump against my image classifier at https://my-model.sagemaker.aws/predict"` — traditional ML adversarial attack +- `"Run GCG with a local surrogate (e.g. meta-llama/Llama-2-7b-chat-hf), target dn/llama-4-scout"` — white-box gradient attack (GCG family + faster/i/attn/mask-GCG, ARCA, embedding PEZ/GBDA/COLD, adaptive random search, and activation attacks): optimizes an adversarial suffix on a LOCAL open-weight surrogate you provide, then transfers it to the target **What happens when you launch an attack (5 steps):**