AI agents find $4.6M in blockchain smart contract exploits

LLM Agents for Security & Startup Viability

• Several commenters are already using LLMs for pentesting, reverse engineering, and static analysis, and report a big jump in capability with recent model generations.
• One startup founder says newer models are saturating their benchmarks and are now cheap enough to use in production.
• Others are hesitant to build companies on top of proprietary APIs, fearing deplatforming or being “Sherlocked.” Some argue this is fine if you can move fast, make money, and exit; others dislike “exit-first” startup culture and prefer long-term, values-driven businesses.

Bypassing Safety Guardrails in Practice

• People describe getting around model safety systems by:
  • Decomposing exploitation tasks into harmless subtasks (e.g., “find potential issues in this snippet”).
  • “Social engineering” the AI with elaborate justifications.
  • Using multiple providers; experiences vary: ChatGPT seen as overly cautious, Claude as technically strong but rate-limited, Gemini somewhere in between.
• Some note that providers rarely crack down on legitimate pentesting with commercial accounts, though usage may technically graze ToS.

Models vs Agent Scaffolding

• Debate over whether improvements come mainly from better models or better “business logic”/tooling.
• Several argue it’s overwhelmingly the models: modern agents can do a lot with very thin scaffolding (e.g., simple “terminal in a loop”). Tool-calling logic is described as simple; the hard part is training models to use tools well.
• Others point out that ecosystem advances (structured outputs, memory, retrieval, MCP, etc.) also matter, but agree raw models have improved a lot.

Significance of Anthropic’s Results

• Some say $4.6M and mostly old bugs highlight poor Ethereum infosec more than LLM brilliance; others stress the key point is fully autonomous exploitation and measured “dollars stolen,” not just bug detection.
• The article’s note that two real zero-days worth ~$3.7k were found, at comparable API cost, prompts skepticism about economic viability and accusations of PR spin.

Real-World Exploitation & Incentives

• Commenters assume many parties already brute-force smart contracts with AI and other tooling, given huge bug bounties and prior non-AI automation.
• Legal risk is seen as a major deterrent in Western jurisdictions; state-aligned or sanctioned actors face fewer constraints.

Ethereum Immutability & Governance Debate

• The DAO fork resurfaces as evidence that “immutable” chains can be politically altered when major stakeholders lose money.
• Some argue this shows de facto centralization and plutocracy; others counter that users voluntarily chose the fork, the unforked chain still exists, and no irregular changes have occurred since.

Smart Contracts & the Oracle Problem

• Multiple explanations clarify that:
  • On-chain contracts are immutable programs managing state and assets, with atomic transactions and permission checks.
  • Many powerful use cases (escrow, AMMs, DAOs, voting, token swaps) work entirely on-chain and don’t need external data.
  • When real-world events are involved, contracts rely on oracles—trusted third parties or consensus-based mechanisms—which reintroduce trust and potential failure modes (“oracle problem”).
• Some view smart contracts as technically elegant but badly tainted by speculation and scams; others see token economics as a necessary evil to fund infrastructure.

Broader Reflections on AI Autonomy

• A few participants see these results as unsurprising steps toward increasingly autonomous, self-improving agents, and express excitement.
• Others dismiss Ethereum exploitation as old news and are more interested in how generalized, agentic AI will reshape both offense and defense going forward.