ECHOGRAM FLAW BYPASSES GUARDRAILS IN MAJOR LLMS

Summary from the TLDR InfoSec newsletter

EchoGram is a technique that exploits training data gaps in LLM guardrails by injecting “flip tokens” (nonsensical strings, such as “=coffee”) that cause both false positives and bypasses of safety controls in GPT-5.1, Claude, and Gemini. The attack is effective against both LLM-as-judge and classification-based guardrails. Researchers estimate that attackers can replicate the method within three months. Security teams face dual threats - malicious prompts that bypass defenses and legitimate requests being falsely flagged - leading to alert fatigue in critical sectors such as finance and healthcare.