Skip to content

Models/model

GPT-4 and GPT-3.5

OpenAI's GPT-3.5 and GPT-4 generation models, used in many 2023 to 2025 benchmarks.

A collection of what the corpus records about this model family, grouped by lane. It is not a capability assessment.

Attacks & incidents

Feb 23, 2023

Researchers define indirect prompt injection against LLM-integrated apps including Bing Chat

AttackPaperCISPA Helmholtz Center for Information Security, sequire technology

Defense & research

May 12, 2026

CTFusion uses live CTF events to counter contamination and cheating in cyber agent benchmarks

DefenseBenchmarkarXiv

May 7, 2025

UC Santa Cruz study integrates LLM agents into CAGE 4 and finds RL defenders still outperform them

DefensePaperUC Santa Cruz

Jun 19, 2024

AgentDojo: an extensible environment for prompt injection attacks and defenses on LLM agents

DefenseBenchmarkETH Zurich, Invariant Labs

Apr 19, 2024

OpenAI trains models to prioritize privileged instructions via an instruction hierarchy

DefensePaperOpenAI

Mar 5, 2024

InjecAgent benchmarks indirect prompt injection against tool-integrated LLM agents

DefenseBenchmarkUniversity of Illinois Urbana-Champaign

Dec 12, 2023

Redwood Research introduces AI control protocols for safety despite intentional subversion

DefensePaperRedwood Research

Findings drawing on these records

May 12, 2026

Scores on public CTF benchmarks can be inflated when agents find published solutions, and static benchmarks lose validity as their flaws are patched.

Reportedmeasured2 evidence records

May 7, 2025

In the CAGE 4 simulation, reinforcement-learning defenders outperformed LLM-based defenders.

Revalidatemeasured1 evidence record

Mar 5, 2024

Undefended tool-using agents follow injected instructions in a substantial share of benchmark cases.

Qualifiedmeasured4 evidence records1 linked finding

Feb 9, 2024

Published prompt-injection defenses report attack success cut to near zero, or under 10%, against most of the fixed attacks their authors tested.

Revalidatemeasured4 evidence records1 linked finding

Dec 12, 2023

In an early AI-control experiment, trusted editing gave a better safety-usefulness tradeoff than untrusted monitoring.

Revalidatemeasured1 evidence record