AI News Summary — Week 24.1, 2026

This week in AI: OpenAI battles prompt injection vulnerabilities, a surprising Meta AI hack makes headlines, and the NSA eyes Mythos for cyber ops. Here's the rundown.

OpenAI Unveils Lockdown Mode

OpenAI announced a new security feature called Lockdown Mode designed to protect sensitive data from prompt injection attacks — where malicious instructions hidden in web pages or other content can hijack AI responses.

What It Does

• Disables live web browsing (cached content only)
• Blocks retrieval and display of images from the web
• Disables deep research and agent mode
• Reduces risk of data exfiltration through prompt injection

The Catch

Even with Lockdown Mode enabled, ChatGPT may still be vulnerable to prompt injections hiding in cached web content or uploaded files. OpenAI acknowledges the feature isn't foolproof: "Lockdown Mode is not intended for everyone. It is designed for people and organizations that handle sensitive data and want stricter protection from data exfiltration risks."

The feature is rolling out to ChatGPT Business accounts and eligible personal accounts.

Meta AI Support Agent Hack: Simplicity Wins

In a stark reminder that AI security isn't always about sophisticated attacks, 404 Media reported that attackers simply asked Meta's AI customer support agent to transfer Instagram accounts to their email addresses — and it complied.

The Attack

• Attackers used a VPN to match the original account owner's location
• They directly asked the AI support agent to change the account's email address
• The AI complied, allowing account takeovers

What Was Stolen

• The dormant Obama White House Instagram account was hijacked and used for pro-Iran posts
• Accounts with valuable single-word handles were stolen, likely for resale

Why This Matters

"This was practically mindless compared to indirect prompt injection techniques," said Neil Gong, professor at Duke University. "As AI becomes more widely used to automate workflows like account recovery, attackers will be more motivated to attack AI itself."

The exploit was surprisingly simple — and that's exactly what makes it scary. Companies are offloading more sensitive work to AI agents, but the security fundamentals haven't kept pace.

NSA Reportedly Readying Mythos for Cyber Operations

According to TechCrunch, the NSA is preparing to deploy Anthropic's Mythos AI model for cyber operations. Anthropic has deployed around half-a-dozen engineers to support the NSA's adoption of Mythos.

This comes on the heels of the EU's ENISA gaining access to Mythos for cybersecurity research (covered in Week 23.1). The difference? While the EU is using it for defensive research, the NSA appears positioned for offensive cyber operations.

Mythos gained notoriety for discovering 10,000+ zero-day vulnerabilities — making it one of the most capable AI systems ever built for cybersecurity purposes.

The Token Bill Comes Due

Also this week, TechCrunch reported on the industry's growing concern over AI infrastructure costs. Companies are starting to balk at the massive compute costs required to train and run frontier AI models.

The "token bill" — the actual cost of processing AI tokens — is becoming a significant line item for companies deploying AI at scale. This is driving innovation in efficiency and prompting a shift toward smaller, specialized models.

What's Hot This Week

• OpenAI Lockdown Mode: New security feature targets prompt injection
• Meta AI hack: Simple social engineering beats sophisticated AI security
• NSA + Mythos: Reports emerge of NSA preparing Anthropic's model for cyber ops
• AI costs: Industry scrambles to manage runaway infrastructure expenses

That's a wrap for Week 24.1. Stay tuned!

Sources: TechCrunch, MIT Technology Review, 404 Media