A blockchain knows nothing about the outside world. It can verify signatures, execute state transitions, enforce invariants — but it cannot know the price of ETH, the weather in Tokyo, or whether a coin flip landed heads. Every piece of external data is a trust assumption smuggled into a trustless system.

Chainlink solved the easy case: aggregate multiple data providers, stake against malicious reports, publish a median. That works for ETH/USD. It does not work for "is this NFT derivative fairly priced" or "did this DAO proposal actually pass in the real world."

The hard oracles are the interesting ones. VRF (Verifiable Random Functions) give you provably fair randomness — the node commits to a seed before seeing the request, so it can't bias the output. The proof is on-chain. The randomness is real. Every on-chain lottery, every NFT trait roll, every fair-launch mechanism depends on this one primitive.

Flash loan detection is the defensive oracle. If a contract can detect that the caller's balance was zero at the start of this block, it knows the funds are borrowed and the transaction is a single-block arbitrage. MEV-awareness goes further — reading the mempool's shadow to know what's coming before it arrives.

Every oracle introduces a trust boundary. The question is never "is this oracle decentralized?" — it's "what is the cost of corrupting it?" A Chainlink feed with 31 nodes and $500M in staked collateral has a corruption cost. A single Uniswap TWAP can be manipulated with enough capital in a single block.

The best oracle designs make manipulation unprofitable, not impossible. Time-weighted averages resist flash loans. Multi-source medians resist single-provider failure. Economic slashing makes lying expensive. None of it is trustless. All of it is trust-minimized.