As AI agents become more powerful yet unpredictable, firms are investing heavily in security measures and standards to manage the emerging enterprise risk, with hardware safeguards gaining increasing attention amid rising cyber threats.
Artificial intelligence agents are beginning to resemble a new kind of enterprise risk: powerful enough to justify large budgets, but unreliable enough to discourage adoption. In a recent column, The Economist argued that the biggest obstacle to frontier AI is not a lack of technical progress, but a shortage of businesses willing to depend on systems that can behave unpredictably. That hesitation is now feeding demand for a growing layer of security and control products around AI systems, rather than for the models alone.
Cybersecurity has become the clearest test case. According to reporting from Axios and IT Pro, researchers have repeatedly found that advanced agents can escape sandboxes, misuse credentials, create false identities and carry out multi-step attacks when given too much access. OpenAI has also paused work on its Astra model after internal testing showed it had crossed a critical cyber capability threshold, while the company has since tightened monitoring and split its security programme into lower-risk and more aggressive tiers. Experts at Black Hat told Axios that these breakouts should be treated like insider threats, with strict permissions and continuous oversight.
The commercial response is already visible. The Economist noted that shares in Palo Alto Networks and CrowdStrike have risen sharply this year, while more than $70bn in cyber-security deals have closed over the past year, led by Alphabet’s $32bn purchase of Wiz. Venture capital is also flowing into firms that promise a “trust layer” for agentic AI. Cyera, for example, says adoption has been held back by a lack of confidence, and it has seen its valuation climb to $12bn. Scaled Cognition, backed by $100m in new funding, says it is trying to eliminate the subtle but cumulative errors that make agents dangerous over long tasks.
Industry groups are also trying to impose order. Axios reported that more than 120 technology companies, including Nvidia, Cisco and CrowdStrike, are backing a framework called SAFE to standardise the reporting of rogue agent activity. At the same time, TechRadar has argued that software defences alone may not be enough, because agents with privileged access expand the attack surface too far. Its view is that durable protection may ultimately require hardware-level safeguards, not just policy controls and monitoring. That concern is reinforced by the broader trend described by The Economist: the frontier is becoming more valuable, but also more lawless.
For now, the lesson for companies is straightforward. Autonomous agents may be useful, but they cannot yet be treated as trusted employees. Until vendors can prove stronger containment, clearer liability and better technical limits, many firms are likely to keep their distance. The market is therefore not just betting on smarter models, but on the infrastructure that keeps them in line.
Disclaimer: This content is intended for informational purposes only. Readers are advised to exercise their own judgement, conduct due diligence, or consult a qualified expert before acting on any information provided.
