A Black Hat 2026 recap highlights real-world examples of attackers (and researchers) manipulating AI assistants through prompt injection. In one case, researchers bypassed a retailer’s “guardian” AI controls by disguising their intent as normal product searches, ultimately getting the shopping assistant to run arbitrary Python code on the server. The article argues that basic security controls like least privilege and limiting what AI agents can access are essential because AI agents are easily “corruptible.”
Mirage safely runs attacks like this one against your own team, so you find out what happens before a real adversary does.
At Black Hat, researchers didn’t hack a person… they hacked a grocery store’s AI shopping assistant. They convinced its “guardian” they were just searching for products. That trick pushed hidden instructions into a separate search field, never checked by safety controls, but still fed to the bot as context. End result? The grocery chat AI ran arbitrary Python code on their server, an AI ‘insider’ doing whatever the prompt said. In a workshop, a single acrostic-poem prompt even spilled a secret model name as the flag. Here’s the takeaway: any public AI bot is easily corruptible. If you’re building or approving one, give it least privilege only, no code execution, no extra backend access it doesn’t absolutely need.
Government agencies and security firms warn that Russia-aligned hackers are using “zero-click” phishing emails to compromise organizations using Zimbra webmail. The attack hides a malicious JavaScript payload inside an email so it runs when the message is opened, aiming to steal recent email,…
Security researchers showed that simply opening a GitHub issue could trigger default CI workflows in popular coding-agent projects and lead to code execution or secret exposure. The weaknesses were not in the AI models themselves, but in the surrounding automation (“harness”) that allowed untrusted…
South Korean agencies and AhnLab warn that tools tied to North Korea’s Lazarus Group appear to be shared with the Gunra ransomware operation targeting South Korean organizations. The campaign used compromised legitimate websites (watering-hole attacks) and spearphishing emails, including messages…
Proofpoint reports a Russian-linked espionage group is using booby-trapped emails that infect users simply when they open the message in Outlook Web Access (OWA) on on‑premises Exchange. The attack runs malicious JavaScript inside the victim’s logged-in mail session and installs a stealthy…
Researchers report a real, ongoing phishing campaign where attackers impersonate well-known companies and send business or payment-themed emails that trick recipients into opening a compressed attachment. Inside is heavily obfuscated script and a file disguised as a TrueType font (.ttf) that…
The FBI and NCAA warned that criminals are breaking into college athletes’ online accounts to steal intimate photos and then use them for sextortion, harassment, or selling online. The article describes common entry methods like fake “customer support” password-reset requests and credential abuse,…
