OpenAI backs cyber defenders with expanded grant program

While headlines keep warning that hackers will weaponize AI to generate malware and scale phishing attacks, what about the folks defending the gates?

OpenAI just shared an update on its Cybersecurity Grant Program, offering a look at what it has funded over the past year. Launched in 2023 to put cutting-edge models into the hands of cyber defenders, the program has pulled in more than 600 applications from researchers and developers around the world.

Hardening models against attacks

A big chunk of the spotlighted work focuses on protecting LLMs themselves from adversarial exploitation.

At UC Berkeley, Professor David Wagner’s lab is building defensive techniques against prompt-injection attacks to make large models harder to hijack. Over at Dartmouth, Professor Adam Breuer’s lab is tackling data reconstruction attacks—where adversaries query a model to extract private training data. Traditional defenses usually kill accuracy and slow down training, but Breuer's team is designing methods that block data leakage without hurting performance.

Meanwhile, Mithril Security rolled out an open-source proof-of-concept for confidential inference. By deploying AI models inside GPU-based secure enclaves powered by Trusted Platform Modules (TPMs), they demonstrated how organizations can send sensitive prompts to AI providers without exposing raw data, even to server administrators.

AI agents on the digital frontline

Other grantees are teaching AI to hunt bugs and fight off real intruders.

Researchers at Boston University are training LLMs to detect and patch code vulnerabilities before malicious actors can exploit them. Startup Coguard uses AI to automatically catch software misconfigurations across networks and clusters, replacing clunky legacy rule-based engines.

At UC Santa Cruz, researchers are designing autonomous defense agents using foundation models, testing how well they triage threats and respond to network intruders compared to traditional reinforcement learning. Over at MIT CSAIL, researchers are exploring how LLM agents handle Capture-the-Flag (CTF) challenges and automate red-teaming decisions using a plan-act-report loop.

On the tooling side, independent grantee Gabriel Bernadett-Shapiro created an AI OSINT workshop and starter kit, training international atrocity crime investigators and intelligence students at Johns Hopkins University. OpenAI also noted that defenders mostly use ChatGPT to translate complex log events into plain English, build log parsers, write analysis scripts, and summarize incidents under tight deadlines. To keep adoption going, OpenAI is extending its free account initiative from ChatGPT Plus to Team and Enterprise plans, kicking off with the Research and Education Network for Uganda (RENU).

Why it matters

We think OpenAI’s grant showcase highlights a refreshing, grounded reality for AI security. It is not about deploying a sentient, self-healing cyber matrix; it is about automating the soul-crushing grunt work that drowns human analysts every single day.

For security engineers and IT teams, using LLMs to parse alien-looking logs or review configuration files is already working and worth adopting right now. The barrier to entry is minimal, and the time saved during an active incident response is very real.

However, the double-edged sword here is impossible to ignore. Every capability that helps a defender spot a zero-day flaw in open-source code can just as easily help an adversary pinpoint an exploitable vulnerability. Furthermore, turning network defense over to autonomous AI agents introduces an entirely new attack surface: if prompt injection remains largely unsolved, relying on an LLM agent to guard network perimeters looks like a massive potential liability waiting to happen.