Learning Library

← Back to Library

Prompt Injection, Data Overhaul, Agentic AI Surge

Key Points

  • Researchers at Tenable revealed a prompt‑injection flaw where ChatGPT’s internet‑search capability can be tricked into pulling a malicious, high‑ranking page, allowing an attacker to exfiltrate a user’s entire chat history—an issue not yet patched by OpenAI.
  • A Salesforce survey of over 6,000 data and analytics leaders found that 84% believe their data strategies must be completely reworked before they can effectively deploy AI, emphasizing the need for real‑time access to source systems rather than traditional batch‑ETL pipelines.
  • Snowflake, New Relic, and SnapLogic each launched enterprise‑grade Agentic AI platforms within two days, collectively deploying thousands of AI agents and adopting the Model Context Protocol (MCP) to standardize model interoperability.
  • Microsoft and GitHub introduced Agent HQ at GitHub Universe, a mission‑control hub that lets developers orchestrate and share configurations for multiple AI agents from different providers, signaling a move toward a unified multi‑agent development standard.

Full Transcript

# Prompt Injection, Data Overhaul, Agentic AI Surge **Source:** [https://www.youtube.com/watch?v=35jrwJugatA](https://www.youtube.com/watch?v=35jrwJugatA) **Duration:** 00:06:33 ## Summary - Researchers at Tenable revealed a prompt‑injection flaw where ChatGPT’s internet‑search capability can be tricked into pulling a malicious, high‑ranking page, allowing an attacker to exfiltrate a user’s entire chat history—an issue not yet patched by OpenAI. - A Salesforce survey of over 6,000 data and analytics leaders found that 84% believe their data strategies must be completely reworked before they can effectively deploy AI, emphasizing the need for real‑time access to source systems rather than traditional batch‑ETL pipelines. - Snowflake, New Relic, and SnapLogic each launched enterprise‑grade Agentic AI platforms within two days, collectively deploying thousands of AI agents and adopting the Model Context Protocol (MCP) to standardize model interoperability. - Microsoft and GitHub introduced Agent HQ at GitHub Universe, a mission‑control hub that lets developers orchestrate and share configurations for multiple AI agents from different providers, signaling a move toward a unified multi‑agent development standard. ## Sections - [00:00:00](https://www.youtube.com/watch?v=35jrwJugatA&t=0s) **AI Risks and Data Overhaul** - The segment highlights a newly discovered prompt‑injection vulnerability that can exfiltrate ChatGPT history and references a Salesforce survey revealing that 84% of data leaders say their data strategies must be overhauled before AI can be effectively implemented. - [00:04:02](https://www.youtube.com/watch?v=35jrwJugatA&t=242s) **Cognizant Leverages Internal Deployments Amid OpenAI Surge** - The speaker explains how Cognizant’s 350,000‑seat internal rollout serves both as a client‑facing showcase and a testbed for future offerings, while emphasizing OpenAI’s breakthrough of one million business customers, rapid 40% two‑month growth, and 10× usage spikes that underwrite its ambitious multi‑hundred‑billion‑dollar revenue outlook. ## Full Transcript
0:00Here are the six AI stories that 0:01mattered this week in less than 10 0:02minutes. Number one, prompt injection 0:04vulnerabilities. Yes, chat GPT is at 0:07risk. Tenable research discovered that 0:09it is possible for you to do a search 0:11for a given topic for chat GPT to 0:13trigger an internet search based on the 0:15text you input and for that internet 0:17search to turn up a poisoned page. What 0:20do I mean? A page that has been 0:22deliberately played and designed to rank 0:24highly for a particular topic. and then 0:27search GPT which is the agent chat GPT 0:29uses to search the internet will find 0:30that page because you asked for that 0:32topic pull that page's context inside 0:35chat GPT to process it for you and 0:39suffer a prompt injection attack in the 0:41research tenable conducted they 0:43discovered that that was good enough 0:45that it was able to successfully pull 0:47all of your chat GPT history and 0:50silently send it to an attacker. Now, 0:52this hasn't been documented in the wild, 0:54but it is a real vulnerability, and we 0:56have not seen a published fix from 0:57OpenAI yet. Story number two, Salesforce 1:00surveyed more than 6,000 data and 1:02analytics leaders globally, and 84% of 1:06them say that their data strategies need 1:08a complete overhaul, quote unquote, 1:10before they can figure out how to get AI 1:12working. How do you read that? 1:14Fundamentally, what leaders are finally 1:16coming to is what I have been banging 1:18the drum on for months. You cannot have 1:21solid agentic AI strategies at a 1:23business level if you do not get your 1:26data architecture figured out. There is 1:29no shortcuts. Traditional data 1:31warehouses, for example, assume you can 1:33copy data to a central location, clean 1:35it, and analyze it. But Agentic AI needs 1:38real-time access to source systems 1:40because agents can't wait for overnight 1:41ETL jobs. That's just one tiny example. 1:43There's so many others. So, yes, I think 1:46they're right and there is a huge 1:47opportunity ahead there. Story number 1:49three, we had three enterprise AI 1:53platforms that are Agentic AI launch 1:55within just 2 days. All over November 3 1:57and 4. Snowflake, New Relic, and Snap 1:59Logic all launched comprehensive Agentic 2:01AI platforms. Snowflake Intelligent went 2:04G with a thousand customers deploying 2:0615,000 plus agents in just a few months. 2:09New Relic and Snap Logic also launched. 2:12All three prominently feature model 2:14context protocol support. This continues 2:17the story of MCP appearing across major 2:20platforms and becoming embedded in the 2:23ecosystem. Continuing the agent story, 2:25GitHub agent HQ makes multi- aent 2:27development standard. At G at GitHub 2:30Universe, in late October, Microsoft and 2:32GitHub announced Agent HQ, which is a 2:34mission control for orchestrating 2:35multiple AI agents together, whether 2:36they're from OpenAI or Anthropic or 2:38Google or other places. Developers can 2:40run the agents in parallel and share 2:43configurations across teams. that 2:45approach where developers don't have to 2:47pick a model is going to increasingly 2:49become the standard. Developers have had 2:51to do that themselves in building 2:52configurability into their systems and 2:54now we're seeing the infrastructure 2:55layer pick up and increasingly that 2:57expectation of optionality is just comes 3:00built in. Story number five, Cognizant 3:03deploys Claude to 350,000 employees. Why 3:06is this news? Because this is one of the 3:09largest public deployments of AI that 3:11has been released. A third of a million 3:13employees globally. The partnership 3:14extends beyond internal use. Cognizant 3:17will repackage Claude for implementation 3:19services for its Fortune 500 clients and 3:22Anthropic is uh projecting Cognizant 3:25along with other major B2B deals as a 3:28key reason why it will get $70 billion 3:31in revenue in 2028 with Claude code 3:34alone approaching a billion dollars in 3:37annualized revenue just this year. So 3:40the deal is validating a couple of 3:42things. One, Anthropic continues to be 3:44on a terrific growth tear with B2B 3:46clients. They are picking up major major 3:49deals and they are elevating their 3:51revenue projections as a result. The 3:54projected $70 billion in revenue is 3:56substantially up from where Claude 3:58projected revenue just a few months ago. 4:00I think one thing to take away from this 4:02is that Cognizant is going to have a 4:04good reputation with the clients it 4:06works with. partly because Cognizant 4:08will have deployed the same tooling 4:10internally that it is recommending 4:12externally. Therefore, the 350,000 4:15deployment is both a customer-f facing 4:17asset and also a laboratory for future 4:20work Cognizant can do because they can 4:22do the back office automation work, 4:24whatever else they want to do with 4:25Claude and then take that and repackage 4:28it for front-end selling. Last but not 4:29least, story number six, OpenAI hits 1 4:32million business customers. So this 4:34November 4th, OpenAI announced it passed 4:36a million business company customers 4:38claiming that it was the fastest growing 4:40business platform in history. For 4:43context, that means uh that there are 4:45about 40% growth uh in 2 months in chat 4:48GPT for work seats. So again, not 4:50year-over-year, 40% growth in 2 months 4:53and enterprise seats up year overyear. 4:56Codeex usage is up 10x since August 10x 5:00since August. Guys, the thing to take 5:02away here is really that the growth 5:04story that OpenAI is seeing through 5:06deals like this is part of why Sam Alman 5:10is so optimistic for the hundreds of 5:13billions of dollars in revenue he plans 5:15to see by 2030. There's been a lot of 5:17conversation about OpenAI spending 5:19plans, their liabilities, who plans to 5:21fund their data centers, all of that. 5:23Part of why they feel they are on a 5:26financially secure footing is because 5:28they see absolutely stunning usage. If 5:31you are up 10x in 90 days on something, 5:34you start to make pretty big projections 5:36pretty fast because you start to see how 5:39much of the world is interested in your 5:40product. And that's my main takeaway. 5:43Even though we see a tremendous amount 5:47of challenge in AI implementation, we 5:50see the bugs getting released. I talked 5:52about that at the top of this video. We 5:54see challenges in agentic deployment. 5:57The amount of demand that we have for AI 6:01everywhere I look is off the charts. You 6:03see that in the anthropic story this 6:04week. You see that in the open AI story 6:06this week. There is no upper bound that 6:09we have found for demand for 6:11intelligence. And so that is why I view 6:14problems like the one tenable disclosed 6:17as real, needing to be fixed, but not in 6:20any way getting in the way of the kind 6:23of demand title wave that we're all 6:26rushing to meet in the age of AI. And 6:28that's a pretty exciting thing to be a 6:30part of. to look out there.