OpenAI jacks up bug bounties to $100K on the road to AGI

When an AI service hits 400 million weekly active users, security stops being a background IT task and turns into an existential mission.

OpenAI just detailed a broad set of cybersecurity initiatives, laying out how it plans to protect its systems, models, and emerging agents on the road toward AGI.

A 5x bounty bump for critical bugs

The most immediate update hits white-hat researchers in their wallets.

OpenAI is raising its maximum bug bounty payout for critical findings from $20,000 to $100,000. It is also rolling out limited-time promotional bonuses for eligible vulnerability disclosures in targeted categories.

Meanwhile, the company is revamping its Cybersecurity Grant Program. Over the past two years, OpenAI evaluated more than 1,000 proposals and funded 28 projects spanning prompt injection, autonomous cyber defenses, and secure code generation.

The updated grant program is now taking pitches across five priority areas: automated software patching, model privacy, threat detection and response, AI security tool integration, and agentic security. It is also introducing microgrants that hand out API credits so researchers can prototype ideas without waiting for long funding reviews.

AI fighting AI, from open source to autonomous agents

Beyond paying external hackers, OpenAI is putting its own models on defense duty.

The lab says its models have achieved top scores on public code security benchmarks, successfully identifying vulnerabilities in open source software. OpenAI plans to responsibly disclose these flaws to relevant project maintainers as it scales the effort.

Internally, the company teamed up with security firm SpecterOps to run continuous adversarial red teaming across its corporate, cloud, and production infrastructure. Following a recent spear phishing campaign aimed at its own staff, OpenAI shared the attackers' tradecraft with competing AI labs to strengthen collective defenses across the sector.

The real test lies with autonomous agents like Operator and deep research. To prevent prompt injection and rogue behaviors, OpenAI is building a unified pipeline and modular framework designed to provide real-time visibility and policy enforcement over agent actions.

Even physical hardware is getting tighter guardrails. For next-generation infrastructure projects like Stargate, OpenAI says it is integrating zero-trust architectures and hardware-backed protections down to the physical supply chain.

Why it matters

For enterprise teams and developers, this is a clear signal that OpenAI wants to make autonomous agents enterprise-ready. Handing system access to tools like Operator requires enormous trust, and higher bounties combined with real-time agent monitoring help lower that adoption hurdle. Still, we think developers should remain cautious: prompt injection remains fundamentally unsolved, and placing sensitive production workflows entirely in an agent's hands still carries real operational risk.

Across the industry, OpenAI's $100,000 bounty cap raises the stakes for competitors like Anthropic and Google. Elite security talent naturally gravitates to the highest payouts, effectively pushing rival labs to match these rewards or risk having critical flaws found and traded on underground markets. OpenAI's decision to share phishing tradecraft also looks like a tacit admission that frontier AI threats are scaling faster than any single company can handle alone.

We think pairing grant funding with automated open source patching is a clever move. It lets OpenAI battle-test its code-auditing models on real-world software while scoring goodwill points with the developer community. The vulnerability, however, is execution speed: as agents gain more desktop and browser autonomy, real-time safety guardrails could quickly become latency bottlenecks, and striking the balance between safety and responsiveness will be an ongoing struggle.