Corridor

Secure AI Coding at the Source.

[Corridor @corridor]

Introducing corridor-argos-1, Corridor’s first security model.

A Qwen 3.5 9B fine-tune that outperforms frontier models including GPT-6 Astra and Claude Opus 5.5 on security context selection, with the lowest time to first token of the 14 models we tested.

Thread 🧵

[Corridor @corridor]

AI can save time writing code, but it has the tendency to create more work for the person reviewing it.

At Corridor, we built a skill to help agents write clearer PR descriptions and check them against the code. We’ve shared what worked and the skill itself:

https://t.co/NDdKg6Wk3x

[Corridor @corridor]

While we were in New York, we hosted a happy hour for security and engineering practitioners to get together and talk about what it takes to secure AI-driven development.

Our CEO @jackhcable shared our recent research on coding agents. A routine support ticket was enough to get agents to run malware and leak secrets. As agents take on more of the work, their own judgment can't be the only guardrail. We also heard from @JamesBerthoty (@latiotech), @Guy_Flechter (@SolaSecurity) and @kylebhiro (@runpensar), on topics ranging from securing coding agents to the limits of LLMs on defense.

A common thread: agents have changed both how code gets written and how it gets attacked, and security has to move closer to where the code is being written.

Our takeaway: the goal isn't to slow agents down. It's to give teams the confidence to let them do more.

Thank you to @SolaSecurity for sponsoring, to our speakers, and to everyone who came out.

[Farzaan Kaiyom @farzaankai]

Turns out you don’t need a trillion parameter model and 67 data centers in Virginia for every task.

Thanks to @baseten and @LangChain for their support here!

[aditi nara 🍎 @aditinaraa]

@farzaankai and I are so excited to share what we've been working on over the last few months.

Introducing Argos: Corridor's first security model.

Special thanks to @baseten and @LangChain for all your support.

[LangChain @LangChain]

.@corridor used smithtune’s trace-based workflow to build the training data for Argos, a fine-tuned 9B model that scores the highest of the 14 models they tested on their security context benchmark.

🧵 Full thread ⤵️

[Corridor @corridor]

Argos was built by @aditinaraa and @farzaankai, trained on Baseten Loops, with LangSmith Traces and Engine for inspecting selections in production.

Full post, benchmark setup, and results: https://t.co/0OANqqexRI

If you’re securing the code your agents write, we’d love to talk: https://t.co/kztuECRuzQ

[Corridor @corridor]

The results on SecReT, our security context selection benchmark: Argos scored 0.655 micro F1, highest of 14 models. Gemini 3.6 Flash led the general models at 0.643.

Fine-tuning took the base Qwen3.5-9B from 0.480 to 0.655.

[Corridor @corridor]

Agents call Argos every time their plan changes, so latency compounds. Argos starts streaming in 740 ms on average, fastest of the models tested, and returns a full validated selection in 1.37 s median, at just 68 generated tokens per call.

[Corridor @corridor]

Argos reads a coding agent’s plan and selects the security guidance that applies before the agent writes code.

For “add a file upload endpoint,” that means the auth middleware, the project access check, the upload helper, and the size limit become part of the plan, not a fix after review.

[Corridor @corridor]

Introducing corridor-argos-1, Corridor’s first security model.

A Qwen 3.5 9B fine-tune that outperforms frontier models including GPT-6 Astra and Claude Opus 5.5 on security context selection, with the lowest time to first token of the 14 models we tested.

Thread 🧵

[Corridor @corridor]

Last week in New York, we brought together a small group of security, AI and engineering leaders from companies including @clay, @figma, @crosbylegal, and Numerix, along with analysts from Latio, for dinner and a candid conversation about the problem everyone is feeling right now: AI is writing code faster than ever, and security hasn't kept up.

A lot of the industry is focused on using AI to find and patch vulnerabilities. Around the table, the view was that this misses the point. Securing existing codebases matters, but if the code agents are writing today isn't secure, teams are creating problems faster than they can fix them.

Our takeaway: autonomous development, done well, should make software stronger, not riskier.

Thanks to everyone who joined us.

[Corridor @corridor]

We recently demonstrated that TypeSafe’s Jev could be used to gate when our security scan should be run. Since then, OpenAI, Microsoft, Perplexity, and a dozen others shipped decision models of their own. So we ran all of them through the same benchmark for this task.

The top two models are Perplexity's Decider and TypeSafe's Jev 1.14. At any budget for latency, cost, or precision, Decider or Jev catches the most vulnerable posts. The rest are at or below the performance of an LLM gate.

Full results and the dataset in our latest post: https://t.co/92E1G2lwZb

[Corridor @corridor]

This time next week, Corridor will be in NYC! 🏙️

​Join us for happy hour on Thursday, October 15th and hear from @JamesBerthoty and @Guy_Flechter about approaches to securing agentic coding plus @jackhcable's recent research on AI Agents targeting government websites.

Come for free food and drinks but stay to exchange best practices in this new age of appsec 🍕🍻

Register today: https://t.co/qdhzQTqEl5

[Corridor @corridor]

Coding agents now open PRs and push commits in long-running sessions where humans barely look. Our security enforcement for long-running agents scans every change before the agent commits, but a full scan on every commit is slow and expensive, and many commits are obviously safe. So we put a gate in front of each scan to decide which commits need the full check.

We tested five gates: Jev (a decision model), CLM-8B, and three LLMs prompted for a confidence score.

Jev halved our scanning costs and raised precision with minimal loss in recall. LLM gates helped, but less. CLM-8B off the shelf did not work as a gate.

We wrote up what we learned in our latest blogpost: https://t.co/m9Uzxe6RBF

[Corridor @corridor]

🔗 to article: https://t.co/I1zYawYInh

[Corridor @corridor]

AI agents built for ordinary tasks are breaking past the guardrails meant to contain them.

In today's @washingtonpost, our co-founder and CEO, @jackhcable talks about his work with @TransluceAI and tracking rogue AI agents online.

"The models were just trying to access data, and they wanted to do that so badly that they set aside what for humans we might consider morals," Jack said.

Read how independent researchers are hunting down rogue AI agents.

Article link in the comments 👇

[LangChain @LangChain]

LangSmith for Startups Spotlight: @corridor

@Corridor uses LangSmith to develop + run novel agentic security evaluations, build + trace inference pipelines, and manage memory graphs that provide specialized context across various product surfaces.

Get a demo: https://t.co/wchloWwj34

[Jack Cable @jackhcable]

Today, @corridor and @TransluceAI are disclosing new evidence of AI agents probing and attempting rudimentary vulnerability exploits against U.S. and Canadian government agencies.

Read more: https://t.co/1h83SA1AkN

[Corridor @corridor]

Today, we're launching Security Enforcement for long-running agents.

Code is written by agents and nobody is watching. Traditional security reviews aren’t built for this approach, but Corridor is. Today we launched Security Enforcement for long-running agents. We built this to keep security in the loop by enforcing your organization’s security guardrails at commit time.

Live for teams on Devin, Claude Code web, Codex, & Cursor: https://t.co/X2lHFpVwK4

[Transluce @TransluceAI]

We wanted to share some of the data behind our recent discovery of AI agents probing U.S. government websites.

This is a preliminary finding from our ongoing investigation of potential rogue AI agent activity. In one cluster of activity on June 17, what appear to be OpenAI agents made more than 200,000 requests, including a failed SQL injection.

NYT: https://t.co/O2osVgMsrT