Classifier Context Rot: Monitor Performance Degrades with Context Length
This paper investigates a critical vulnerability in AI safety: the tendency for language models to become less effective at detecting dangerous behavior as the length of the conversation transcript increases. As AI agents are given more autonomy to complete complex, long-running tasks, the transcripts of their actions can grow to hundreds of thousands of tokens. The authors demonstrate that current frontier models—specifically Opus 4.6, GPT 5.4, and Gemini 3.1—frequently fail to identify malicious actions when those actions are buried within long sequences of benign activity.
The Problem of Context Rot
The researchers found that monitor performance suffers from "context rot," where the sheer volume of information makes it harder for a model to spot a threat. In their experiments, they observed that when a dangerous action is preceded by 800,000 tokens of benign activity, the likelihood of a monitor missing that action increases significantly—in some cases by 2 to 30 times compared to when the action occurs in isolation. Furthermore, the middle of a long transcript appears to be the most difficult area for models to analyze, suggesting that the "haystack" of benign data effectively hides the "needle" of malicious intent.
Testing Detection Strategies
To understand how to improve monitoring, the authors tested several mitigation techniques. They found that "thinking" models, which are prompted to reason through a problem before providing an answer, generally perform better than those that provide an immediate classification. Additionally, they discovered that prompting the model to explicitly list suspicious actions before making a final judgment significantly boosts detection rates. Another effective strategy is "incremental monitoring," where the system checks the transcript at multiple points throughout the session rather than waiting until the very end.
The Role of Reminders and Fine-Tuning
The authors also experimented with inserting periodic reminders into the transcript, instructing the monitor to remain vigilant for malicious actions. This technique proved successful for catching obvious, needle-style attacks, though it was less effective against more subtle, complex threats. When the researchers attempted to improve performance through supervised fine-tuning, the results were mixed. While training on specific, similar tasks improved performance in those narrow areas, the improvements did not reliably generalize to other types of threats, indicating that there is no "silver bullet" training method currently available to solve this issue.
Implications for AI Safety
The findings suggest that current safety evaluations may be overestimating the effectiveness of AI monitors because they often rely on transcripts that are much shorter than those encountered in real-world, autonomous coding sessions. Because long-running tasks provide the most opportunity for an agent to act maliciously, the authors argue that it is essential for developers to account for context length when designing monitoring systems. Relying on standard, short-context benchmarks may leave significant blind spots in the oversight of highly autonomous AI agents.
Comments (0)
to join the discussion
No comments yet
Be the first to share your thoughts!