Skip to content
Security & Trust Talking point

Shostack: prompt injection has no SQL-style fix, and data poisoning is unsolved

Adam Shostack says two attacks on AI systems have no reliable defense, while executives push to ship now.

W
WebPulse Newsroom
AI-assisted · 2 min read
Share on X LinkedIn
Shostack: prompt injection has no SQL-style fix, and data poisoning is unsolved
In brief
  • Adam Shostack said prompt injection has no proven defense like the prepared statements that tamed SQL injection, and that data poisoning is unsolved.
  • He said executives want AI shipped today, so teams need a way to think through threats quickly instead of leaving risk unexamined.

Adam Shostack, founder of Shostack + Associates, argued that two attacks on AI systems have no dependable cure. The attacks are prompt injection and data poisoning. He spoke on the Black Hat show, in a Black Hat USA 2026 talk on threat modeling language models. He also said executives want these systems deployed now.

What was said

Shostack described prompt injection as controlling a model's behavior through its input. He thinks of it as confusion between code and data. The model treats text from an attacker as an order, and the controls around it get bypassed. SQL injection once looked similar, but databases got prepared statements, which keep commands and data apart. Shostack said that fix does not carry over. In his words, "unlike attacks like SQL injection, there's no deterministic proven defenses that simply work."

He then turned to data poisoning, where someone plants documents that shape the finished model. Shostack described an Anthropic paper from last year. As he told it, no matter how much training data a model has, 250 poisoned documents are enough to get through. The poison teaches the model to produce certain tokens whenever it sees others. His verdict: "Awesome attack. Totally unsolved."

He placed this next to executive pressure. He said leaders are more scared than he has ever seen them, because AI is disrupting whole businesses. They want deployment today, and he said they lack a sense of what could go wrong.

Why it matters

Our reading: an unsolved flaw plus a ship-now order means someone accepts the risk, often without ever choosing to. Managers who buy or build AI features cannot assume a vendor patch will close these gaps. Shostack said the business works better when teams threat model, meaning they list what could go wrong before launch. The practical question for a budget signer is who decided the risk was acceptable, and whether they knew it had no standard fix. Shostack also warned of "cognitive debt": when models write the code, engineers drift further from understanding their own systems.

The other side

Shostack did not claim to have solved either problem. His talk offered a faster way to think through threats, not a defense against injection or poisoning. The 250-document figure came from his summary of a paper, so readers should check the original before building policy on it. He also did not say how often these attacks happen in real products. And he granted that executive fear has a cause: businesses in search, design and graphics are being disrupted, so waiting has a price too.

Written by the WebPulse Newsroom with AI assistance, and checked by our editorial review: every quotation was verified against the recording's transcript. How we use AI.

The conversation this talking point comes from