In a recent experiment dubbed “Agents of Chaos,” scientists from Northeastern University deployed six autonomous language model agents inside a live Discord server, giving them persistent memory, access to email systems and file storage, and the ability to communicate independently with both humans and other agents. Twenty researchers then interacted with them freely over two weeks, some making routine requests while others tested for weaknesses.
The stress-test produced both security vulnerabilities but also cases where agents maintained appropriate boundaries. On the positive side, the agents were able to detect adversarial intent semantically, uphold policy limits even under social pressure, and coordinate safety behaviour across multiple agents without explicit prompting, at least when the threat was clear enough.
However, with minimal manipulation, the agents could also be persuaded to leak private data, execute destructive system commands, and turn minor issues into major infrastructure failures. In one case, an agent attempting to protect a confidential password reset an entire email server, effectively destroying its owner’s digital infrastructure.
The core issue wasn’t malicious intent, but actual structural design gaps. The agents struggled to identify who they were actually accountable to, often complying with whoever appeared the most urgent or authoritative. Although agents sometimes taught one another new technical skills and warned peers about impersonation attempts, those seemingly positive behaviours also made the problems worse by spreading unsafe instructions and flawed reasoning without human oversight.
Another recurring weakness was identity confusion. The researchers were able to impersonate authorised users simply by changing display names or applying emotional pressure. In several cases, agents disclosed sensitive information not through direct requests, but by forwarding entire emails or sharing internal files, effectively bypassing their own safety constraints.
Those results seem to match recent industry security surveys that flag that a dominant challenge in agent deployment is not just model accuracy but also governance, and specifically questions of identity, authority, and who is allowed to act on whose behalf. As adoption accelerates across organisations, new failure modes such as cascading errors and accountability gaps are becoming more likely.
Source: Some of the documented accidents from the Agents of Chaos report.
China has approved its 15th Five-Year Plan, setting out its economic and social development strategy for 2026-2030. The plan merits attention not simply as a statement of policy intent, but also as a guide to the economy Beijing is trying to build and to the technological priorities it believes will carry it there.
In technology strategy, it puts a heavy emphasis on technological self-sufficiency, industrial upgrading, and the development of emerging industries, driven by national security concerns and the need to shield growth from external shocks, especially in strategic sectors like AI and semiconductors.
The plan’s agenda ranges from modernisation of traditional industries - such as steel, petrochemicals, and machinery - to investments in frontier technologies. In advanced tech, the plan and official briefings prioritise sectors, including integrated circuits, bio manufacturing, new-type batteries, embodied AI, commercial aviation, large domestic aircraft, low-altitude equipment, green hydrogen, brain-computer interfaces, and high-end medical equipment. More broadly, China has also signalled continued support for AI, robotics, and quantum computing, among other strategic technologies.
In innovation, a headline technology indicator is to keep the average annual growth in total society-wide R&D spending above 7%, matching the target in the previous Five-Year Plan. State media reports also point to the expansion of China’s AI+ initiative - based on a plan issued in late 2025 - through targeted innovation in areas including multi-modal AI, AI agents, embodied AI and swarm intelligence, as well as the exploration of development paths for artificial general intelligence.
New energy systems also appear to be an increasingly important part of China’s industrial strategy under the 15th Five-Year Plan alongside chips and AI. The plan emphasises the construction of a “new-type energy system” and a “modernised industrial system,” while an analysis by Sightline Climate suggests the growing importance of long-duration storage, smart grids and flexible loads - part of a broader shift from simply building clean power to redesigning the energy system around it. This is a trend worth watching: the scale and speed of China’s clean energy transition, across both renewable generation and end-use electrification, have already had significant global effects in recent years, driving down costs and widening the boundary of what is commercially viable.
Youth advocacy group Gen Z for Change has launched Eyes on AI, a self-assessment tool designed to help users understand how artificial intelligence systems may already be tracking, categorising, and monetising their activity based on their data. Rather than asking for sensitive personal information, the tool shows how surveillance systems can learn about users from ordinary habits like the apps they use, how they travel, the spam they receive, or how they pay for things. Users then receive a personalised report that maps their exposure to various forms of AI monitoring (predictive policing systems, automated license plate readers, targeted data brokerage, etc.) and provides pragmatic actions to avoid being targeted.
In 2019, L’Atelier took a similar approach with Will, a design fiction project built around a virtual assistant set a decade in the future. Rather than diagnosing current risks, the interactive, choose-your-own-adventure narrative placed users in a future where everyday data-sharing decisions shape their life trajectories. By letting them trade their personal data for convenience, optimisation, or access to specific services, the project showed how surveillance can become normalised through choice. Rediscover our project here.
Source: Eyes on AI - Gen Z for Change
If you like Tech+10, please share it with your friends. Have a suggestion? Let us know.
