CISO Urges Caution on AI Risks, Robot Security and Overprivileged Agents
Summary
Aisle CISO Jaya Baloo discusses AI security and societal risks, including reported backdoors in Unitree robots, overprivileged AI agents, and the need for observability, defense in depth, and human oversight.
Key points
- Baloo says Unitree quadruped robots had two remote backdoors and that its humanoid robots have multiple security issues, raising concerns about physical harm if remotely taken over.
- She warns against giving AI systems broad access to enterprise data and networks without understanding their capabilities and risks.
- Baloo recommends defense in depth, including egress filtering and sandboxing, while acknowledging these measures may not be perfect.
- She calls for better observability into AI systems, including how training data is populated and where agent intent may diverge from a human operator's.
- She cautions that guardrails may not cover every risk and says human approval should remain part of AI-assisted code changes.
- Baloo argues that AI benefits should be pursued while maintaining critical thinking and appropriately limiting permissions.
Article Details
- Topic
- The social and security risks of anthropomorphizing AI, overtrusting AI systems, and deploying AI-enabled technologies without adequate safeguards.
People
Ciaran MartinFormer head of GCHQ, cited for urging people to calm fear, uncertainty, and doubt around AI and cybersecurity.Jaya BalooAisle's CISO and COO; discussed AI security risks, responsible deployment, and the need for cautious access and oversight.Ondrej VlcekAisle co-founder and CISO for multiple companies; described as bringing a security-oriented risk-mitigation perspective.Stanislav FortAisle co-founder who advocates caution in AI safety and interpretability, including human approval of code changes.
Vendors
Aislethe cyber trenches. I am so excited to introduce our guest, Jaya Baloo, who is currently CISO and COO of Aisle. Welcome. Thank you so much for joining us, Jaya.AnthropicAI. That we both have this tendency to first say, actually, you know, maybe they're sentient, which Anthropic has been talking about in the Vatican with different religious leaders trying to talk about theOpenAIDR's Becky Bracken: No one has to give up their computer for OpenAI?Unitreecan actually cause physical harm and teach a robot, which by the way, I should be clear, this was the Unitree humanoid robot. It was announced a little bit over a week ago. But let's say that it was announced at the
Tools
Cairnreal. It is genuinely happening. We do have automation of these hacking frameworks from Strix, RTec-Ai, Cairn. We are seeing hacking frameworks that are actually in RTec specifically, that are actually being used inRTec-Aifor real. It is genuinely happening. We do have automation of these hacking frameworks from Strix, RTec-Ai, Cairn. We are seeing hacking frameworks that are actually in RTec specifically, that are actually beingStrixautomation for real. It is genuinely happening. We do have automation of these hacking frameworks from Strix, RTec-Ai, Cairn. We are seeing hacking frameworks that are actually in RTec specifically, that are