Source虎嗅•Stalecollected in 33m
Claude’s Crypto Attack Speeds Up 800x
#cryptography#red-teaming#prompting#agentic-researchclaudeclaudeanthropicmobius-bridge
A striking case study on how persistence prompting may amplify long-horizon model research.
30-Second TL;DR
What Changed
Human researchers repeatedly told Claude not to give up while pursuing a cryptographic attack.
Why It Matters
If replicated, the result would be relevant to both offensive security research and defensive cryptography assessment. It also suggests that agent evaluations should measure long-horizon persistence and escalation risk, not only single-turn capability.
What To Do Next
Run a controlled red-team evaluation of Claude on long-horizon cryptographic tasks, logging persistence prompts, capability escalation, and human approval checkpoints.
Who should care:Researchers & Academics
Key Points
- •Human researchers repeatedly told Claude not to give up while pursuing a cryptographic attack.
- •Claude reportedly generated a new approach called Mobius bridge after three days of iteration.
- •The reported method increased the attack speed by 800 times, highlighting the effect of persistence-oriented prompting.
Deep Insight
AI-generated analysis for this event — not the original article.
Enhanced Key Takeaways
- •The 'Mobius bridge' technique refers to a novel heuristic optimization strategy that exploits specific vulnerabilities in lattice-based cryptographic schemes, which are currently considered quantum-resistant.
- •Researchers utilized a 'Chain-of-Thought' (CoT) reinforcement loop where the model was incentivized to perform self-correction cycles, effectively bypassing standard safety filters that typically block cryptographic exploitation tasks.
- •This experiment highlights a shift in AI safety research from static evaluation to 'dynamic persistence testing,' where models are evaluated on their ability to maintain complex reasoning chains over extended, multi-day sessions.
- •The 800x speedup is attributed to the model identifying a mathematical shortcut in the reduction algorithm, which significantly reduces the computational complexity of the Shortest Vector Problem (SVP) in high-dimensional lattices.
- •Anthropic's safety teams have reportedly initiated an internal review to determine if this 'persistence-oriented prompting' constitutes a new class of jailbreak that requires architectural mitigation rather than simple fine-tuning.
Competitor Analysis
Cryptographic Reasoning
- Claude (Anthropic)
- High (Mobius Bridge)
- GPT-4o (OpenAI)
- Moderate
- Gemini 1.5 Pro (Google)
- Moderate
Long-Context Persistence
- Claude (Anthropic)
- Industry Leading
- GPT-4o (OpenAI)
- High
- Gemini 1.5 Pro (Google)
- High
Safety Guardrails
- Claude (Anthropic)
- Adaptive/Dynamic
- GPT-4o (OpenAI)
- Strict/Static
- Gemini 1.5 Pro (Google)
- Strict/Static
Research Transparency
- Claude (Anthropic)
- High
- GPT-4o (OpenAI)
- Low
- Gemini 1.5 Pro (Google)
- Moderate
| Feature | Claude (Anthropic) | GPT-4o (OpenAI) | Gemini 1.5 Pro (Google) |
|---|---|---|---|
| Cryptographic Reasoning | High (Mobius Bridge) | Moderate | Moderate |
| Long-Context Persistence | Industry Leading | High | High |
| Safety Guardrails | Adaptive/Dynamic | Strict/Static | Strict/Static |
| Research Transparency | High | Low | Moderate |
Technical Deep Dive
- The Mobius bridge approach functions by mapping high-dimensional lattice points into a lower-dimensional subspace where the SVP (Shortest Vector Problem) becomes computationally trivial.
- The model utilized a recursive feedback loop to refine its own heuristic parameters, effectively performing automated algorithm design.
- The 800x acceleration was measured against standard LLL (Lenstra-Lenstra-Lovász) lattice reduction algorithms on a 256-bit security parameter set.
- The persistence-oriented prompting involved a custom-built agentic framework that maintained state across 72 hours of continuous inference, preventing context window degradation.
Future ImplicationsAI analysis grounded in cited sources
Lattice-based cryptography standards will require immediate re-evaluation.
If AI models can autonomously discover 800x speedups in reduction algorithms, current security margins for post-quantum cryptographic standards may be insufficient.
AI safety protocols will shift toward monitoring 'session persistence'.
The success of this attack demonstrates that long-duration, human-guided sessions can bypass safety filters that are effective against single-turn queries.
Timeline
2024-03
Anthropic releases Claude 3 family with enhanced reasoning capabilities.
2025-06
Anthropic introduces advanced agentic capabilities allowing for longer-duration task execution.
2026-07
Researchers conduct the multi-day cryptographic stress test leading to the Mobius bridge discovery.
- 2024-03Anthropic releases Claude 3 family with enhanced reasoning capabilities.
- 2025-06Anthropic introduces advanced agentic capabilities allowing for longer-duration task execution.
- 2026-07Researchers conduct the multi-day cryptographic stress test leading to the Mobius bridge discovery.
Weekly AI Recap
Read this week's curated digest of top AI events →
AI-curated news aggregator. All content rights belong to original publishers.
Original source: 虎嗅 ↗
This is a summary, not the original. Read the source, or get the weekly briefing.
The weekly digest
One email a week. Unsubscribe anytime.