Paul Price

We partnered with Mozilla to test Claude's ability to find security vulnerabilities in Firefox. Opus 4.6 found 22 vulnerabilities in just two weeks. Of these, 14 were high-severity, representing a fifth of all high-severity bugs Mozilla remediated in 2025.

It is hard to communicate how much bug bounty has changed due to AI in the last 2 months: not gradually and over time in the "progress as usual" way, but specifically this last December. There are a number of asterisks but imo coding agents basically didn't work for security research before December and basically work since - the models have significantly higher quality, long-term coherence and tenacity and they can power through large and long hacking tasks, well past enough that it is extremely disruptive to the default bug bounty workflow. Just to give an example, over the weekend I pointed Claude Code at a new program's scope and wrote: "Here are the target domains. Enumerate subdomains, grab all the JavaScript bundles, run the full analysis pipeline (endpoints, secrets, source-sink tracing, postMessage handlers), fuzz the discovered paths, spider the authenticated surface, check for IDORs on user APIs, test any interesting GraphQL endpoints, and write up an HTML report of everything you find." The agent went off for ~30 minutes, ran into multiple issues (auth failures, WAF blocks, malformed responses), researched solutions, resolved them one by one, analyzed the JS, fuzzed endpoints, tested access controls, and came back with the report. Two confirmed vulnerabilities and a handful of interesting leads. I didn't touch anything. All of this could easily have been a full weekend of manual work just 3 months ago but today it's something you kick off and forget about for 30 minutes. As a result, bug bounty hunting is becoming unrecognizable. You're not manually clicking through Burp Suite and hand-testing parameters one by one like the way things were since this industry started, that era is over. You're spinning up AI agents, giving them targets *in English* and managing and reviewing their output in parallel. The biggest prize is in figuring out how you can keep ascending the layers of abstraction to set up long-running orchestrator agents with all the right skills, memory and instructions that productively manage multiple parallel hacking instances for you. The leverage achievable via top tier "agentic engineering" for security research feels very high right now. My friends and I have been building out custom skill libraries for Claude Code - things like JS static analysis pipelines, authenticated fuzzing, IDOR testing frameworks, GraphQL introspection - and sharing them with each other. Each person's agent gets better as the collective skill set grows. We're finding more bugs in a week than we used to find in a month. It's not perfect, it needs high-level direction, judgement, hacker intuition, oversight, iteration and hints and ideas. It works a lot better in some scenarios than others (e.g. especially for targets with thick JavaScript clients where you can verify findings with a curl command). The key is to build intuition to decompose the target just right to hand off the recon and testing parts that work and help out around the edges with the creative exploitation. But imo, this is nowhere near "business as usual" time in bug bounty.

Tonight, we reached an agreement with the Department of War to deploy our models in their classified network. In all of our interactions, the DoW displayed a deep respect for safety and a desire to partner to achieve the best possible outcome. AI safety and wide distribution of benefits are the core of our mission. Two of our most important safety principles are prohibitions on domestic mass surveillance and human responsibility for the use of force, including for autonomous weapon systems. The DoW agrees with these principles, reflects them in law and policy, and we put them into our agreement. We also will build technical safeguards to ensure our models behave as they should, which the DoW also wanted. We will deploy FDEs to help with our models and to ensure their safety, we will deploy on cloud networks only. We are asking the DoW to offer these same terms to all AI companies, which in our opinion we think everyone should be willing to accept. We have expressed our strong desire to see things de-escalate away from legal and governmental actions and towards reasonable agreements. We remain committed to serve all of humanity as best we can. The world is a complicated, messy, and sometimes dangerous place.

If you seriously think this is gonna find some live complex vulnerability, you’re tripping lmao.

ONS - "Annual average regular earnings growth was 7.2% for the public sector and 3.4% for the private sector"

*JAPAN 4Q GDP GROWS 0.1% Q/Q; EST. 0.4%