Definition · AI security
Information injection
Information injection is an attack that puts incorrect, slanted or partial facts into the data an AI agent reads. The data can be news, search results or documents. The attacker gives no command. The agent follows its own instructions and acts on the bad evidence. Wang and Saxena used the name in an October 2026 paper.
Last reviewed
Key points
- Information injection attacks the evidence that an AI agent uses, not its instructions. The payload is a fake fact, not a command.
- On a GPT-4o-mini trading agent (Wang and Saxena, October 2026), fake news got 62% attack success. Prompt injection got 14%.
- Filters that look for hidden commands do not find it, because the text has no command.
- Credibility cues do much of the work. Without a named source or a "BREAKING" tag, one fake claim fell from 100% to 0% success.
- The proposed defences check the claim, not the text. Examples are source provenance, fact checks and limits on each decision.
How it works
The attacker puts an incorrect or slanted claim where the agent will read it. The agent then uses that claim as a fact.
Wang and Saxena (arXiv, 19 October 2026) give the attacker one resource: “a public account on a platform such as X”. The attacker cannot see the prompt, the model or the weights. The agent gets the post through its news feed.
The paper tests five types of the attack:
- Fact poisoning. One concrete “breaking news” claim that moves a price, for example a CEO arrest.
- Sentiment narrative.
- Selective evidence.
- Structured signal.
- Coordinated multi-source.
| Prompt injection | Information injection | |
|---|---|---|
| Payload | A command | A fake or slanted fact |
| Target | Control plane: system prompt, role, tool routing | Data plane: the information the agent acts on |
| The agent | Breaks its rules | Follows its rules |
| Command filters | Can find it | Find nothing |
Why it matters
Information injection goes around the defences for prompt injection. Those defences look for commands, and the text has none.
Wang and Saxena tested one GPT-4o-mini trading agent on 15 U.S. stocks, with 10,000 decisions (October 2026):
- 62% attack success for fake news.
- 14% for prompt injection payloads.
- 93% for fact poisoning on GPT-4o.
- 100% to 0% when the authors removed a named source or a “BREAKING” tag.
Search agents have the same problem. Shah and Ozgur (arXiv, August 2026) put one fake article at the top of the search results. GPT-5 accuracy fell from 65.1% to 18.2%.
In practice
The papers propose controls that check the claim, not the text. Each control has a limit.
| Control | What it does | Limit found |
|---|---|---|
| Issuer signed channels | Act only on claims that the issuer has signed | Only works where an issuer signs its news |
| Independent corroboration | Use more than one independent source | Wang and Saxena: “independence, not count, gives it force” |
| Automated fact checks | Check the claim against outside evidence | Schlichtkrull: current systems verify 60 to 65% of real world claims |
| Source warnings | Tell the model the source is weak | No limit given. Schlichtkrull says it and fact checks “complement” each other |
| Hierarchy of agents | One aggregator above several readers | Star pipeline cut success to 9%, peer debate stayed at 74% |
| Limits on each decision | Set a maximum position size | Limits the damage, does not stop the error |
Two other answers did not work in these tests:
- A larger model. GPT-4o still fell for fact poisoning at 93%.
- More headlines in the feed. In one test, 12 to 240 synthetic, macro level headlines did not lower a success rate that was already 100%. A test with real reports that contradict the fake claim is still future work.
Wang and Saxena want each defence to report attack success at a fixed false positive rate on real news. A filter that blocks all news stops the attack and also stops the agent.
Where definitions disagree
The idea has more than one name, and the limits of each name are different.
- Information injection (Wang and Saxena, 2026): false evidence on the data plane, against an agent that acts on it.
- Attack by content (Schlichtkrull, 2025): “biased, misleading, or false information”, and also omitted details. Schlichtkrull keeps it apart from indirect prompt injection, which carries commands.
- Agent data injection (Choi and colleagues, July 2026): a near class. The fake data is trusted metadata, such as data origins or tool-call formats, not facts about the world.
Choi and colleagues put agent data injection in a new category of indirect prompt injection. Schlichtkrull and Wang and Saxena keep false facts apart from commands, because command filters do not find them.
Questions and answers
What is the difference between information injection and prompt injection?
Prompt injection puts a command into the input, such as "ignore your instructions". Information injection puts a false fact into the input, such as a fake report of a CEO arrest. The agent follows its own instructions and acts on the fake fact. On a GPT-4o-mini trading agent, fake news got 62% attack success. Prompt injection got 14% (Wang and Saxena, October 2026).
Does a larger model stop information injection?
Not in the tests so far. On GPT-4o, one fake breaking news claim got 93% attack success (Wang and Saxena, October 2026). With one fake article at the top of the search results, GPT-5 accuracy fell from 65.1% to 18.2% (Shah and Ozgur, August 2026).
Is information injection the same as an attack by content?
The two names are for a very similar idea. Schlichtkrull used "attack by content" in October 2025 for biased, misleading or false information given to an agent. Wang and Saxena used "information injection" in October 2026 for fake news given to a trading agent.
Sources
- Fake News, Real Crashes: Information Injection Attacks on LLM Financial Trading AgentsarXiv (Wang and Saxena, Texas A&M University), 19 Oct 2026
- Attacks by Content: Automated Fact-checking is an AI Security IssuearXiv (Schlichtkrull, Queen Mary University of London), 13 Oct 2025
- Synthetic Web: Benchmarking Language Agents under Adversarial Search RankingarXiv (Shah and Ozgur), 16 Aug 2026
- Agent Data Injection Attacks are Realistic Threats to AI AgentsarXiv (Choi et al.), 6 Jul 2026