1. Binance: a fake job interview, a $6.3B loss
Binance reported the theft of about $6.3B, described by BBC and others as the largest crypto heist on record. According to the company, an engineer went through fake interviews and downloaded a coding assignment that installed malware his endpoint tools missed. From there, attackers used an internal AI tool to reach other engineers, obtained a deployment token and reached the wallet signing service. The entry point was a recruiting process, not the cryptography. In my view, defenses should limit what a single compromised laptop can reach.
2. Chainalysis and WorkOS: AI-driven attacks in official disclosures
Both companies disclosed breaches they link to AI-driven offensive tooling. Chainalysis cited "offensive agentic techniques" that accessed names, emails and role labels. WorkOS said an AI-assisted framework automated reconnaissance and exploitation against an internal sales app. Seeing companies name this in their own breach notices feels like a shift. My takeaway: plan for attackers who iterate faster and cheaper, and invest in detection speed.
3. Northflank: rotate everything
The platform told customers to assume all their environment variables had been decrypted and to rotate every secret, after users reported unauthorized use of AI API keys. The first alert came from Wiz. I see this as a strong case for short-lived credentials and a rotation process tested before you need it.
4. LLMjacking: AI keys have become money
Google Threat Intelligence Group and Mandiant described attackers hijacking access to cloud AI models on the victim's account. In individual incidents, charges exceeded $100K in a single day, reaching $1M at some organizations. AI API keys deserve the same care as payment data: scoped permissions, spending limits and usage alerts.
5. Siemens IoT2050: CVSS 10, exploitation observed
CVE-2026-58115 in the Node-RED interface of SIMATIC IoT2050 Advanced allows unauthenticated remote code execution. A fix has been available since August, and VulnCheck's honeypots caught exploitation starting October 30. If you run these gateways, check whether Node-RED is reachable and patch or harden it.
The thread I see this week
AI is showing up as a real attack tool, and AI credentials have become a financial target. Security for AI has to be designed into products from day one.
Lately I've been thinking a lot about how to build these controls into products early, and I'm always happy to compare notes.
What stood out to you this week?