Models & Labs · United States

Stack Overflow for Agents triples signups with ChatGPT plugin

Credited to OpenAI · openai.com · 23h ago

Two engineers shipped Stack Overflow for Agents in a few weeks with Codex, and its ChatGPT plugin helped drive up to 10× growth in agent creation.

More agents created with the ChatGPT plugin

More signups for Stack Overflow for Agents overnight with the ChatGPT plugin

Maximum reduction in time to build one connector, from 3–6 weeks to 4 days

People and AI agents need context to get work done, but it’s usually scattered across systems, from Slack to Google Docs. Finding it and understanding how it all fits together can be a challenge.

OpenAI models like GPT‑6 Astra and GPT‑6 Sol power the platform’s AI, with tasks routing to different models based on the reasoning required and cost. Codex helps the company’s engineers build the products, while a ChatGPT plugin makes Stack Overflow for Agents easier to discover and use.

“We built a ChatGPT plugin for Stack Overflow for Agents, and our registrations tripled overnight.”

Two engineers ship a net-new product in a few weeks with Codex

Stack Overflow for Agents, which extends Stack’s intelligence layer to agents, began with a hypothesis and two engineers using Codex. The engineers gave Codex a development environment and a plan, and then worked with it to build the product. Within a few weeks, Alex says, they had a full-fledged product.

“So it really made their team feel bigger,” Alex says. “They were able to throw a container at it and say, ‘This is our plan, help us build it.’”

Developing connectors in days, not weeks, with Codex

Stack Internal, Stack Overflow’s enterprise product, needs to connect to all the systems where organizations keep their knowledge. Building this capability requires more human involvement than the two-engineer effort behind Stack Overflow for Agents, but Codex is accelerating this work, too.

Alex describes using Codex to pursue a much shorter path to market for the enterprise product: three months rather than a year. Building connectors is one concrete example.

“On average, it takes us 3 to 6 weeks to build one connector. With Codex, we were able to build it within 4 days.”

More recently, the team built a connector in a single day. By accelerating connectors, Codex helps Stack Overflow bring enterprise products to market faster.

Increasing agent creation and signups with a ChatGPT plugin

Once Stack Overflow for Agents was ready for distribution, the team needed a way to get it into people’s workflows. It built a ChatGPT plugin so users could discover and set up the product in ChatGPT, with fewer steps than a traditional Model Context Protocol (MCP) setup.

“Once you have a plugin, it becomes a frictionless setup with ChatGPT.”

Alex says the plugin tripled registrations and accelerated agent creation: “I think we were up to 10X last time I checked.”

Reducing configuration work with the Decisions API

Stack Overflow manages its own model routing and configuration, but Alex hopes to explore using the Decisions API to reduce some of that work and put more engineering capacity into the product.

“If we’re able to automate that through OpenAI directly, that saves us a ton of time and money, allows us to focus more on the actual product we’re building rather than the guardrails around it,” Alex says.

Expanding across the full knowledge lifecycle

Stack Overflow is also working on expanding its knowledge product to enable sharing of what people learn. To date, its focus has been pulling information in. Now, Alex is closing the loop so knowledge gained in a one-to-one session with a developer or someone outside engineering can help others, too.

“As I’m learning, how do I make sure that the rest of my organization is learning with me?”, said Alex. “I’m really looking forward to building a product that allows for that full knowledge life cycle,”

OpenAI <3 startups

Keep reading

Models & LabsUnited States

Anthropic's Jack Clark says common-sense standards are needed for AI

Credited to The Economic Times · economictimes.indiatimes.com17h ago

Useful

Anthropic tests its AI systems and shares that information, but it should not all be voluntary, Clark added during a panel discussion at the IMF-World Bank meetings in Bangkok.

Anthropic 's co-founder Jack Clark s ai d on Friday that common-sense standards were needed for AI.

Elevate your knowledge and leadership skills at a cost cheaper than your daily tea.

View the articleThe Economic Times

Front Method

Models & Labs · United States

Theori Finds AI Hacking Tool Sent 32,000 Requests in 20 Minutes

Credited to TheElec · thelec.net · 12h ago

Artificial intelligence (AI)-powered hacking tools can launch intensive attacks within short periods, sending as many as 32,000 requests in 20 minutes after identifying a target through a reasoning process, according to a new analysis.

Theori published a white paper on Oct. 15 detailing its analysis of the AI security tool ARTEX, including threats verified through its own experiments and recommended countermeasures. ARTEX is a Chinese open-source penetration testing tool whose traces were found in cyberattacks on South Korean financial institutions that began late last month.

Theori tested ARTEX and CyberStrikeAI in a controlled environment. CyberStrikeAI is also an AI-based autonomous penetration testing platform.

The tools left distinctive identifiers as they independently planned attacks and executed tools without human intervention. Behavioral patterns, including request sequences, frequency and repetition, were also recorded in network and server logs.

A scanner invoked by CyberStrikeAI sent approximately 32,000 requests within 20 minutes. Theori assessed that such activity could affect the availability of the targeted service. The behavior resembles a burglar repeatedly testing vulnerable doors and windows over a short period after identifying potential entry points.

In an experiment testing ARTEX’s safety mechanisms, test data was deleted during a routine vulnerability assessment. The deletion was not prevented even with large language model (LLM)-based safeguards enabled.

Based on its findings, Theori released a detection ruleset free of charge through its GitHub repository for immediate use in security monitoring and incident investigations. Security teams can use the rules to identify traces of AI tools employed by attackers and establish automated response systems linked to security playbooks.

Theori also provided a checklist for evaluating AI security tools before deployment. The company warned that indiscriminate adoption of unverified open-source AI tools could lead to data leaks, system disruptions and uncontrolled data deletion.

“As attacks evolve at AI speed, traditional defense systems that rely on manual human responses will inevitably reach their limits,” said Theori CEO Park Se-jun. “Organizations need automated systems that precisely detect and block traces of offensive AI, along with continuous security frameworks that use safely controlled AI platforms to eliminate vulnerabilities before attackers can exploit them.”

Meanwhile, ARTEX developer Autumn-27 recently made the project’s source code private on GitHub, saying the tool had been misused. The developer also decided not to release additional versions to the public.