Twitxy Saved

SnowCrash Labs

@SnowCrashLabs

Joined November 2025
10 Following 10 Followers 206 Posts

Latest posts 5

SnowCrash Labs @SnowCrashLabs
@Google's threat intel team found an attacker pipeline that diffed security patches, wrote exploits and ran them with no human in the loop. A first, GTIG says.

The patch window is now a race against an agent. Measure yours.

cloud.google.com/blog/topics/th…
cloud.google.com GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold | Google Cloud Blog This AI threat update provides GTIG’s findings on adversarial misuse of AI including Gemini and other non-Google tools.
SnowCrash Labs @SnowCrashLabs
@OpenAI's ChatGPT Wallet lets agents pay after a biometric check on each purchase. Later, OpenAI expects users to pre-approve by spending limit or category.

Then the control is the agent staying inside the limit. Test that before you delegate.

openai.com/index/introduc…
openai.com Introducing ChatGPT Wallet Introducing ChatGPT Wallet, a secure way for agents in ChatGPT Work, Codex, and dots to complete purchases with user approval and existing card protections.
SnowCrash Labs @SnowCrashLabs
A sandbagging audit scored 0.978 AUC. After two rounds of adaptive retraining: 0.560. The hidden trigger still erased 40-49 points of math accuracy.

Small open-weight models, new @arxiv paper. Ask evaluators for adaptive-attack results.

arxiv.org/abs/2611.14950
arxiv.org Stress-Testing Sandbagging Detection Under Adaptive Attack Developers, auditors and researchers read evaluation scores as evidence of what a model can do. A model that sandbags, scoring low on purpose while keeping the capability, turns a low score into...
SnowCrash Labs @SnowCrashLabs
414,210 sign-ins under a popular AI agent's name hit one marketplace in a week: credential stuffing, per Cequence via Help Net Security. Real agents looked like browsers.

An agent's name proves nothing. Judge behavior.

helpnetsecurity.com/2026/11/12/bot…
www.helpnetsecurity.com Bots pose as AI agents, AI agents pose as humans, and defenses believe both - Help Net Security AI agent impersonation hid 414,210 credential stuffing visits at one marketplace, while a real agent passed itself off as a desktop browser.