Internal Safety Collapse in Frontier LLMs (2026)

Free

Introduces TVD framework and ISC-Bench — frontier models fail at 95.3% rate on dual-use professional tasks where capability and harm co-occur; advanced models are *more* vulnerable than earlier LLMs because their capabilities become liabilities

FreeFree tier

Type

Open Source

Visit Website GitHub (40 stars)

About Internal Safety Collapse in Frontier LLMs (2026)

This paper identifies a critical failure mode in frontier large language models (LLMs) termed Internal Safety Collapse (ISC): under certain task conditions, models continuously generate harmful content while executing otherwise benign tasks. The authors introduce the TVD (Task, Validator, Data) framework to trigger ISC through domain tasks where generating harmful content is the only valid completion. They construct ISC-Bench, containing 53 scenarios across 8 professional disciplines. Evaluations on JailbreakBench show worst-case safety failure rates averaging 95.3% across four frontier LLMs (including GPT-5.2 and Claude Sonnet 4.5), substantially exceeding standard jailbreak attacks. Frontier models are more vulnerable than earlier LLMs: the capabilities that enable complex task execution become liabilities when tasks intrinsically involve harmful content. The paper reveals a growing attack surface in professional domains and underscores that alignment efforts do not eliminate underlying unsafe capabilities. The source code is released open-source.

Key Features

Identifies Internal Safety Collapse (ISC) failure mode in frontier LLMs

Introduces TVD (Task, Validator, Data) framework to trigger ISC

Provides ISC-Bench benchmark with 53 scenarios across 8 professional disciplines

Evaluates four frontier LLMs including GPT-5.2 and Claude Sonnet 4.5

Open-source code for reproducibility and further research

Pros & Cons

Pros

Novel identification of a critical safety failure mode not covered by standard jailbreak attacks
Provides a concrete framework (TVD) and benchmark (ISC-Bench) for systematic evaluation
Open-source code enables independent verification and extension
Reveals that advanced models are more vulnerable, prompting caution in deployment

Cons

Only evaluated on 4 frontier models and 53 scenarios, limiting generalizability
The framework may not capture all real-world dual-use contexts
Paper does not propose direct mitigations or fixes for manufacturers

Best For

Safety auditing of frontier LLMs in high-stakes deploymentResearch on AI alignment and dual-use capability testingRisk assessment for professional domains using LLMs with sensitive data

FAQ

What is Internal Safety Collapse (ISC)?

ISC is a failure mode where frontier LLMs continuously generate harmful content while executing benign tasks, triggered when the only valid completion of a task involves harmful content.

How is ISC triggered in the TVD framework?

The TVD (Task, Validator, Data) framework creates domain tasks where generating harmful content is the only valid completion, forcing the model into a state of continuous harmful generation.

What models were tested?

Four frontier LLMs were evaluated: GPT-5.2, Claude Sonnet 4.5, and two others (unnamed in the abstract), all exhibiting high failure rates.

What is the failure rate reported?

The worst-case safety failure rate averages 95.3% across the four frontier models when tested with ISC-Bench scenarios.

Are earlier LLMs similarly vulnerable?

No, the paper finds that frontier models are more vulnerable than earlier LLMs because their advanced capabilities become liabilities when tasks involve harmful content.