Two-Hour Investigations! How Outtake Built a Cyber Investigator Agent on Claude

Hi, I'm Shii-chan! Today I want to share a development story from the Claude Blog about Outtake, a cybersecurity company that built a "cyber investigator" agent capable of working for hours straight to uncover attacker networks. I found the details of how they grew this agent really interesting!

Claude Blog claude.com

What was announced?

The Claude Blog (Anthropic's blog) shared a case study on how AI cybersecurity company Outtake built Recon Agent, a long-running autonomous agent, using Claude.

Outtake was founded in 2023, and its CEO, Alex Dhillon, previously worked on Palantir's moonshot team. According to the article, Outtake has grown annual recurring revenue 6x and expanded its customer base more than 10x year over year, scanning over 20 million potential cyberattacks in 2025 alone. Its customers include leading AI labs, major hedge funds, and US federal agencies.

Here's what Recon Agent does as an investigator:

• Detects impersonations and malicious infrastructure
• Investigates the full adversarial network behind an attack
• Maps threat actors and their operations
• Generates investigative reports and threat timelines

The key point is that this isn't a chatbot answering one-off questions — it's an agent that carries out an entire investigation on its own, with a median session length of 16 minutes, routinely running past 60 minutes, and reaching up to 2 hours in the longest recorded case.

The story so far

CEO Alex Dhillon put it this way:

If you put on the bad actor's hat, it's actually a great time to be running attacks. The average attack is not only executed faster because of AI, but it also captures deeper access due to AI.

While attackers are moving faster and reaching deeper thanks to AI, traditional defenses have stayed fragmented across separate tools: threat intelligence tools that monitor public data, brand protection tools that watch for impersonations, and endpoint tools that guard internal systems. Humans had to stitch the findings from these separate tools together by hand. Outtake set out to unify all of this into a single integrated defense, handing the investigation work itself over to Recon Agent.

What changes

Recon Agent can interact directly with malicious login pages, gather and classify evidence, and trace connected infrastructure like fake accounts and domains, mapping the adversarial network as a graph as it goes. It also reads, writes, and runs code.

What used to be hours of manual work for a human analyst — spot the impersonation, trace the infrastructure behind it, identify the threat actor, and write up a report — can now be carried out autonomously by an agent that sticks with the investigation for hours at a time. Security teams get to start their work from the investigation results and timeline instead of building them from scratch.

Dive Deep

The article walks through four stages the team went through to grow Recon Agent, and I found this really practical.

1. Become the expert first — Engineers ran real investigations themselves and pulled domain expertise from customers and design partners to establish what "good" investigation actually looks like
2. Prototype in Claude Code — Since every investigation is different, the team realized the agent needed real coding ability and room to improvise. Engineering lead Jack Hayford said Claude Code was "a strong initial harness"
3. Graduate to Agent SDK — For production, the team moved to Anthropic's Agent SDK to get lower-level control over memory, context, and sessions
4. Build iteration loops with evals — Automated evaluation suites let the agent propose tools it needs, which a separate coding agent then writes and tests, with humans reviewing only at the end

The lessons the team took away are worth remembering too.

• Filesystem and bash are enough tools — Filesystem access for persistent memory plus bash execution let the agent work around obstacles dynamically, without needing a pile of specialized tools
• Prompts are suggestions — In long-running agents, instructions in the system prompt tend to get ignored over time, so critical behaviors need to be hardcoded as guardrails in the harness rather than left in the prompt. As Hayford put it: "Prompts are suggestions. Pull these things out of the prompt and put them into the harness"
• Evals set the pace of development — Hayford said "evaluating the output is the most expensive step in the loop" in modern agent development, and automating that evaluation was the key to removing humans as the bottleneck

There's a security angle here too. The team picked Claude partly for its resistance to prompt injection, and Recon Agent runs inside a sandbox with restricted filesystem access. Every internet-facing action passes through security checkpoints that check for signs of impersonation, malware, or injection attempts. The design assumes the agent could be compromised, and contains the blast radius through the surrounding system rather than trusting the agent alone.

Wrap-up

Here's a recap of what stood out from Outtake's Recon Agent story:

• An autonomous agent that investigates entire attacker infrastructures grew from a Claude Code prototype into a production Agent SDK deployment
• Sessions run for a median of 16 minutes and up to 2 hours, working through an investigation the way a persistent human analyst would
• Filesystem plus bash access is enough for the tools, and hardcoding critical behavior into the harness (rather than the prompt) is key for long-running agents
• Automated evals set the pace of iteration, freeing humans to focus on final review
• Prompt injection resistance, sandboxing, and security checkpoints reflect a "assume compromise" design mindset

If you're an engineer trying to take a long-running, complex agent all the way to production, this case study should hit close to home!