返回资讯中心
外部精选
软件工程
#开发者实时流

研究显示二元逻辑削弱AI安全防护

来源标题:Binary Logic Breaks AI Safety: "Yes, No, Probably"

一项研究指出,将AI安全判断压缩为「是/否」的二元逻辑会削弱模型对模糊或边界情况的识别能力,引入「可能」等概率判断有助于提升安全性。研究认为当前安全对齐在二元框架下存在盲区。

Hacker News 最新kate83822026/11/14 · 3 天前12 分钟阅读

以下正文同步自 Hacker News 最新,版权归原站所有,已转换为易读排版。

LLM

Ai Architecture

Prompt Injection

Cybersecurity

Binary Logic Breaks AI Safety — and Is Analyzing Prompt Injections Pointless? “Yes, No, Probably!”

Ecaterina Sevciuc

4 min read·

Just now

--

Imagine a simple everyday situation. Someone asks you: “Want some coffee?” You reply: “Yes, no, probably” (“Да нет, наверное”).

To a Russian speaker, this answer is completely natural — it carries a subtle mix of mild hesitation, mood, and context. Now, how would a non-native speaker unfamiliar with the nuances of the language interpret this? Most likely, they would ask a clarifying question, because in their logic, “yes” and “no” simply cannot coexist at the same time — and let’s not even get started on the “probably” thrown into the mix.

Now imagine we forced an AI system to deal with this exact phrase.

The industry has conditioned us to believe that a machine only has two possible decisions: execute or refuse. And that is a dead end. Why? Because without asking clarifying questions, this response turns into pure chaos for a binary algorithm. Is it consent or rejection? Black or white?

So what is the AI supposed to do? And what should it actually be doing in situations like this?

We took Large Language Models — systems that are inherently probabilistic, fluid, contextual, and deeply human-like — and tried to squeeze them into the rigid binary logic of traditional software: 0 or 1, Safe or Unsafe, Allow or Block.

We trained models designed to communicate with humans on strict mathematical logic, and then unleashed that setup into the real world. And now we act genuinely surprised by “Black Octobers,” bypassed guardrails, and bizarre agent behavior in production. When you try to force multidimensional human language into the Procrustean bed of binary rules, the system inevitably starts cracking at the seams.

This exact same binary trap shows up in how the industry argues about AI security architecture. One camp argues that all the noise around prompt injections is just empty hype, claiming that real security lives exclusively in classic infrastructure, APIs, IAM, and access controls. The other camp tries to patch holes with brittle, surface-level input filters.

This division is artificial. Arguing over what matters more — the lock on the front door or the thickness of the walls — loses all meaning when an attacker walks right through the window using a legitimate key. By definition, an AI agent is authorized to call backend endpoints on a user’s behalf. If the model falls for a behavioral manipulation, traditional infrastructure scanners won’t flag any technical anomaly — to the backend, the request looks 100% legitimate.

In the era of autonomous agents and multi-step tool execution, the rules of the game have changed:

Consider a simple agent tasked with managing user accounts. A user submits the instruction:

From a traditional binary perspective, the request looks completely legitimate. The user is authorized. No malware is involved. No obvious security rule is violated.

Yet the command remains dangerously ambiguous:

A binary system must either execute or refuse. A more resilient system introduces a third option:

“Please clarify what you mean by inactive users.”

This additional question is not merely a usability feature — it becomes a security control. The clarification step transforms uncertainty into explicit intent before any irreversible action takes place.

The solution is straightforward: we need to stop panicking at the sight of ambiguity and give our systems the right to hesitate.

Instead of a primitive True/False kill switch, we need to move toward weighted risk scoring (assigning calibrated trigger weights, such as a 0.3 threshold for borderline cases) and clarifying dialogue logic (clarify). This is the exact approach I am currently building into the architecture of my open-source framework, AURA — where benign yet ambiguous scenarios are routed into a dedicated contour, allowing the system to clarify the user’s real intent rather than blocking them at the door.

A security layer shouldn’t jump the gun just because a prompt contains a “suspicious” context or a familiar trigger keyword. A smart agent should know how to pause, ask clarifying questions, resolve the uncertainty, and respond accurately to actual human intent — just like a human does when they hear: ”Yes, no, probably.”

In other words, ambiguity is not a bug to eliminate. It is information to process.

Otherwise, we will just keep wondering why living human language keeps breaking such logical and understandable (to some))) mathematical schemas.

正文由 探How 从来源站点 RSS 同步,内容未经改写;遇到排版缺失或需要图片、视频时请以原文为准。

打开原文
去社区讨论 →