TechLetters ☕️ Do export controls on quantum computer technology make sense? DeepSeek models are not invading privacy. Voice cloning a minister of defence. Censored AI models.
Is this the time for export controls on quantum computing technology?
Security
Example bad take on DeepSeek to be wary of. Weird uninformed take about DeepSeek LLM (local) models by some 'cybersecurity companies'. They claim that non-executable files with numbers (model weighs) are capable of 'sending telemetry'. Well, nope.
Anthropic considered defending LLMs from jailbreaks. By adding a “general” classifier layer, a model with some basic reference that checks input/output. Experienced infosec folks know this approach well. Now there are two models an adversary needs to jailbreak.
Red Teaming risks in generative AI products. Link.
Malicious open source models are being uploaded to popular repository hugging face. This will be a completely new cybersecurity risk. Now it's merely code execution. But expect tainted/poisoned weights impacting outputs. Python reverse shell script connects to IP 107.173.7.141 on port 4243 and enables remote command execution. On Linux, it spawns a `/bin/sh` shell, on Windows, it launches PowerShell and enables bidirectional communication. To avoid such risks use SafeTensor.
What did the malicious code do? This Python reverse shell script connects to IP 107.173.7.141 on port 4243, enabling remote command execution. On Unix/Linux, it spawns a `/bin/sh` shell, while on Windows, it launches PowerShell and facilitates bidirectional communication. It persistently attempts connection, granting unauthorized remote access and control.
Technology Policy
AI/LLM models intended to help with programming refuse to work. When they encounter a forbidden word (gender, sex, trans and so on). It's important to develop methods to circumvent such truncations. Is it really reasonable to call such methods as 'abuse'?
U.S. Congress will consider the Decoupling America’s AI Capabilities from China Act. Proposing 20 years in jail or $1M fines for individuals and $100M for companies downloading Chinese AI models like DeepSeek. It would also ban working with or investing in Chinese-controlled AI entities.
Other
Voice cloning (AI) used to impersonate Italy's defence minister. In a scam targeting a number of rich people under the pretext of raising millions for a fake ransom for journalists. Among the targets, the CEO of Pirelli, Giorgio Armani, the co-founder of Prada, the owner of Tod's and others. At least one businessman transferred €1 million.
In case you feel it's worth it to forward this content further:
Subscribed
If you’d like to share: