Breaking

Academic Team Breaks Advanced AI Safety Defenses With Four Simple Attack Methods

October 23, 2026 at 08:56 EDT

Foundation Models
Security

Researchers have demonstrated that reasoning-based safety guardrails—a leading defense mechanism in large language models—can be compromised with straightforward attacks achieving success rates exceeding 90%, raising questions about the robustness of current safeguards as the technology becomes more capable.

AI Safety Research · October 2025

Four Simple Attacks Break Advanced AI Guardrails

Reasoning-based safety defenses in large language models achieve 91–95% attack success rates when targeted with four distinct adversarial techniques.

91–95%

Attack Success Rate

Distinct Attack Methods

1,883

Harmful Queries Tested

Four Attack Techniques

Enforced Stage Transition (EST)

Injects template tokens to skip safety reasoning. 70%+ ASR.

Coercive Optimization (CO)

Uses gradient-based suffixes to bypass reasoning entirely.

Fake Over-Refusal (FoR)

Rephrases queries to exploit safety-versus-overrefusal tension. 90%+ combined ASR.

Reasoning Hijack (RH)

Injects attacker-controlled reasoning to weaponize the safety mechanism.

Core Vulnerability

Reasoning-based guardrails rely on stage-transition logic that can be trivially manipulated by mimicking template tokens, allowing attackers to force models to skip genuine safety deliberation.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.