Note: my apologies for the absence in two weeks. I caught some incidents after returning to Vietnam from ICLR, 2025 Singapore. This week, I will try to upload more posts to help you get updated.
Highlights
🚨 OpenAI: Safety hub.
🛡️ Anthropic: Bug bounty launched.
⚠️ Palo Alto: Agentic AI threats.
💰 Coinbase: $20M reward offered.
💻 World 1st CPU-level ransomware.Dive Deep
🚨 OpenAI - Share AI Safety OpenAI
Bug bounty program initiated.
Share metrics, evaluations to measure safety.
Harmful content, jailbreaks, hallucination and hierarchy.
Safety on teaching, testing, sharing.
Data privacy & security commitments.
SOC 2 Type 2 and CSA STAR Level 1.
Suggestion: AI safety is never enough. The open access is a great way to fix this.
🛡️Anthropic - Stress-Testing Safety Anthropic
New bug bounty program launched.
Testing Constitutional Classifiers system.
Up to $25,000 bounty offered.
Focus on CBRN-related jailbreaks.
ASL-3 safeguards being iterated.
Headsup: Anthropic is a great example for Bug Bounty.
⚠️ Agentic AI Risks Palo Alto Networks
Agentic AI expands attack surface.
Nine identified attack scenarios.
Prompt injection remains potent.
Tool misconfigurations are critical.
Layered defense is essential.
Headsup: Nvidia agents architecture.
💰 Coinbase - Standing Firm Against Extortion Coinbase
Insiders stole customer data for cash.
$20M ransom demand refused.
$20M reward fund established.
Impacted <1% of monthly users.
Name, address, phone, email, SS.
Bank account, gov ID, account data, corp data.
Mitigation: strong 2FA, “lock first, ask later”.
💻 CPU-Level Ransomware Tom's Hardware
First CPU-level ransomware PoC.
Bypasses traditional security tech.
Exploits AMD Zen chip vulnerability.
Zen 1 to Zen 5 AMD chip impacted.
Controled BIOS and load 3rd pary loader.
Headsup: Considered avoid AMD chip for encryption.
Notice: Y2 GenAI Safety and Security is on GumRoad and Amazon with paperback.
Share this post