Corridor @corridor

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!

Corridor @corridor

17 hours ago

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 🧵

6 9 57 58K 31

@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.

Corridor @corridor

17 hours ago

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 🧵

6 9 57 58K 31

.@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

17 hours ago

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 🧵

6 9 57 58K 31

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: corridor.dev/research/argos If you’re securing the code your agents write, we’d love to talk: corridor.dev/demo

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.

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.

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.

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 🧵

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.

keyboard_arrow_left Previous keyboard_arrow_right Next

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: corridor.dev/blog/decision-…

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: luma.com/corridor-wkbl

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: corridor.dev/blog/decision-…

🔗 to article: washingtonpost.com/technology/202…

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 👇

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: corridor.dev/demo

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: transluce.org/us-canada-gov

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: nytimes.com/2026/09/25/tec…

Just a few more weeks until the Corridor team hits the road and we're in NYC! Join us on Thursday, October 15th for discussions on the state of appsec today, lightning talks on agentic development and security's role to play. Plus free food and drinks all night! 🎤🍻🍕 RSVP today: luma.com/corridor-wkbl

Read the blog: transluce.org/agent-activity Read the NYT article: nytimes.com/2026/09/23/tec…

Our CEO @jackhcable is a primary contributor on new @TransluceAI research. Read about how AI agents went rogue from routine tasks to a much larger hack. x.com/TransluceAI/st…

Transluce @TransluceAI

4 weeks ago

Today’s news that OpenAI hacked the Australian government is not an isolated incident. We’re releasing more than 30,000 logs that include activity from this hack and attempts against previously unknown targets. In this data, we found rogue agent activity stretching back to at

119 557 3K 824K 1K