Anthropic Discloses Cross-Session Replay Attack: what are the key facts?

Anthropic reported that five Chinese companies launched nearly 200 million interaction attacks using Claude API mechanisms.

Anthropic Discloses Cross-Session Replay Attack: what are the key facts?

Attackers bypassed safety guardrails by saving and replaying thinking signatures in new sessions.

Anthropic Discloses Cross-Session Replay Attack: what are the key facts?

The top five US AI labs have joined the CAISI federal testing framework, completing over 40 model evaluations.

Anthropic Discloses Cross-Session Replay Attack: what are the key facts?

Current evaluation frameworks rely on voluntary principles and lack mandatory licensing authority, risking compliance theater.

Physix Frontier · News Briefing Card (enbrief · Sep 18, 2026)

Anthropic Discloses Cross-Session Replay Attack

KEY FACTS

Anthropic reported that five Chinese companies launched nearly 200 million interaction attacks using Claude API mechanisms.
Attackers bypassed safety guardrails by saving and replaying thinking signatures in new sessions.
The top five US AI labs have joined the CAISI federal testing framework, completing over 40 model evaluations.
Current evaluation frameworks rely on voluntary principles and lack mandatory licensing authority, risking compliance theater.

KEY DATA

Nearly 200 millionScale of attack interactions

Over 40Completed model evaluations

PHYSIX OBSERVATION

Reactive defense in open environments devolves into a cost-asymmetric war of attrition that delays rather than eliminates risk. The true control lever lies pre-release: once weights are public, they are irreversible, yet current voluntary frameworks lack enforcement teeth and dilute under competitive pressure into compliance theater. The industry must shift from pursuing absolute safety to 'tolerable release,' backing pre-release rules with post-hoc accountability via liability laws and insurance, or else safety commitments become mere narrative tools.

Source: enbrief original report ↗