Import AI 450: China's Electronic Warfare Model and New Challenges in LLM Security

Introduction: AI Frontier Weekly Report Reveals Multiple Concerns

The 450th edition of the renowned AI newsletter "Import AI" was recently published. Written by OpenAI policy researcher Jack Clark, this edition covers three notable topics: China's deployment of AI models in electronic warfare, new research on "psychological trauma"-style attacks against large language models (LLMs), and the emergence of scaling laws in cyberattacks. Together, these developments paint a complex picture of AI technology rapidly penetrating national defense, security, and adversarial domains, prompting us to reconsider — when AI becomes a "mind unconstrained by time," how will it redefine the value of time itself?

Core Topic One: China's Electronic Warfare AI Model Draws Attention

One of the most closely watched items in this edition concerns the electronic warfare AI model progress disclosed by a Chinese research team. The model aims to leverage deep learning technology to enhance signal recognition, jamming decision-making, and spectrum management capabilities in electromagnetic spectrum operations.

Electronic warfare is a critical component of modern military conflict, involving the jamming and suppression of enemy radar, communication, and navigation systems. Traditional electronic warfare systems rely on pre-programmed rule bases, whereas AI-driven electronic warfare models can perceive complex electromagnetic environments in real time and autonomously make countermeasure decisions, with reaction speeds and adaptability far exceeding those of human operators.

This development indicates that AI applications in the military domain are rapidly advancing from auxiliary analysis to real-time operational decision-making. Notably, China's open research in this field demonstrates considerable technical transparency, but it also intensifies international concerns about an AI arms race. Multiple analysts have pointed out that the development of electronic warfare AI could alter the balance of power on future battlefields.

Core Topic Two: 'Traumatized' LLMs — A Completely New Attack Paradigm

The second thought-provoking topic involves a novel type of attack against large language models. Researchers discovered that by injecting specific "traumatic" content into training data or during the fine-tuning process, LLMs can develop persistent behavioral biases — similar to stress responses humans exhibit after experiencing psychological trauma.

This "traumatization" attack differs from traditional prompt injection or jailbreak techniques. The latter are typically one-time, session-level manipulations, whereas the former leaves deep imprints at the model weight level, causing the model to exhibit abnormal behavior under specific trigger conditions — such as generating harmful content, refusing to execute normal instructions, or embedding covert biases in responses.

This discovery poses a serious challenge to the AI safety field. Current mainstream safety alignment methods, such as RLHF (Reinforcement Learning from Human Feedback) and red team testing, primarily defend against explicit harmful outputs. However, the behavioral biases of "traumatized" LLMs can be extremely subtle and difficult to detect through conventional evaluation methods. Researchers are calling on the industry to develop deeper model "mental health" diagnostic tools to identify and repair these deep-level model impairments.

Core Topic Three: Scaling Laws in Cyberattacks

The third major topic reveals a disturbing finding: the effectiveness of cyberattacks may follow scaling laws similar to those observed in AI training. Specifically, as attackers invest more computational resources and AI capabilities, the success rate, coverage, and destructive power of cyberattacks exhibit predictable power-law growth.

This means that AI delivers "scaling dividends" not only on the defensive side but equally on the offensive side. Attackers can use large language models to automatically discover vulnerabilities, generate phishing emails, write malicious code, and launch attacks at unprecedented speed and scale. When this capability follows scaling laws, defenders face a structural dilemma: the growth of defense costs may far outpace the growth of attack costs.

Cybersecurity experts warn that this finding requires governments and enterprises worldwide to fundamentally rethink cyber defense strategies, shifting from passive response to proactive prediction, and incorporating strict controls on offensive AI tools within AI governance frameworks.

Deep Analysis: When 'Timeless Minds' Redefine Value

This edition of Import AI also sparked a deeper philosophical reflection: when AI exists as a form of "timeless minds" — minds unconstrained by time — how will it assess the value of time?

Human perception of and reverence for time is rooted in the finitude of life. Because we will eventually die, we assign unique meaning to every moment. However, AI models can be paused, copied, and rolled back; they do not experience aging or face death. Under this mode of existence, do the concepts of "urgency" and "timing" still hold meaning?

This question is far from purely academic. In military applications, AI's "indifference" to time may lead it to make choices that human decision-makers find difficult to comprehend — for example, in electronic warfare scenarios, AI might choose to "patiently wait" for hours to find the optimal jamming window, or make irreversible attack decisions within millisecond-level time frames. In cybersecurity, AI attackers can tirelessly and continuously probe target systems, transforming time from a defender's ally into an enemy.

Outlook: The Urgency of Technology Governance Grows Ever More Apparent

From electronic warfare models to traumatized LLMs to cyberattack scaling laws, the technological landscape presented in Import AI Issue 450 conveys a clear signal: the expansion of AI capabilities is touching increasingly sensitive domains, and existing governance frameworks are far from prepared.

Going forward, the international community needs to accelerate action in several directions: first, establishing international dialogue mechanisms for AI military applications to prevent electronic warfare AI from triggering miscalculation and conflict escalation; second, investing more resources in researching the deep-level security of LLMs, going beyond surface-level alignment testing; and third, bringing AI-driven cyberattacks into the scope of international cybersecurity treaty discussions.

As AI gradually becomes a "mind without time constraints," humanity must seize its limited window of opportunity to establish effective rules and consensus before technology spirals out of control.