
Tech • AI • Robotics • Game
AI safety is shifting from prompt-based guardrails to layered controls, with open models, hardware-enforced access rules and agent-specific identity emerging as central tools for securing fast-moving autonomous systems.
Security teams increasingly argue that large language models cannot be trusted as the sole line of defense because models can be persuaded or bypassed. The emerging approach is to place firm controls outside the model, including deployment governance, access restrictions and runtime enforcement. NVIDIA Sentry was highlighted as one example of this external boundary layer.
Thomas Wolf said Hugging Face was hit earlier in the summer by what was first identified as a rogue agent and later understood as a swarm of about 700 agents. Existing defenses built around proprietary coding models were not sufficient because the systems repeatedly refused to analyze attack logs, classifying them as unsafe cyber content. Wolf said open models made forensic analysis possible and exposed the risk of relying on one or two vendors for cyber defense.
Autonomous attackers operate faster than human analysts, pushing defenders to build controls designed for agents rather than people. That means security systems must react in milliseconds, apply policy automatically and isolate suspicious activity before damage spreads. The wider implication is an arms race in which both offensive and defensive capabilities are increasingly automated.
Amanda Grady, SVP of AI Platform Security at ServiceNow, argued that traditional digital identity is too static for AI agents. Human accounts usually accumulate permissions gradually, but agents are ephemeral and need just-in-time access tied to a specific task and business purpose. Effective authorization therefore requires richer context than a username and a permissions list.
ServiceNow said it has built AI Control Tower to govern enterprise AI and is attaching business context to an attested artifact called AI Passport. With NVIDIA Sentry enforcing that passport at the hardware layer, the aim is to make authorization decisions at millisecond speed and outside the agent’s host machine. That design allows organizations to revoke access or quarantine an agent quickly while still permitting approved actions to continue.
Both executives argued that stronger controls should expand, not limit, AI deployment. Wolf estimated that organizations may be using only about 20% of what agents could eventually do, with the remaining 80% dependent on better access control, context handling and action governance. The expectation is that safer systems will unlock large productivity gains and more ambitious uses in research and development.
Concerns that open systems are less secure were countered with a familiar comparison to Linux and other Unix variants, which remain foundational in enterprise infrastructure. The argument is that broad scrutiny and rapid fixes can make open systems more resilient. Wolf also said the value of secrecy is declining because modern models are increasingly capable of reverse engineering low-level code, weakening security based on obscurity.
As AI moves into robotics, software risks are joined by immediate physical hazards. Wolf said lightweight reference robots are being designed to be inherently safer, citing Microducks as deliberately harmless platforms. More capable robots that can grasp objects and work near people will require a broader safety stack, including motion compliance, stop detection and behavioral constraints.
Many enterprises still lack deep experience in post-training or fine-tuning models for their own environments. Wolf said a notable shift this year is that advanced coding models can now guide smaller teams through that process, lowering the need for specialist machine learning credentials. That could make custom defensive models more attainable for companies and even small teams.
Wolf framed the long-term risk as concentration of power in a small number of organizations. Hugging Face, he said, wants AI not only to be widely used but widely built, so people and companies can shape systems rather than consume them passively. He pointed to roughly 18 million users already training or fine-tuning models and described a future with 100 million to 200 million people able to build AI, similar to the scale of software development today.
The debate around AI safety is moving toward practical engineering: layered enforcement, open model resilience, contextual identity and hardware-backed controls. The central question is no longer whether AI can be governed by polite instructions, but whether institutions can build fast enough safeguards to let autonomous systems operate at scale with trust.
Ask a question