Generative AI

Almaty, Kazakhstan, 28.04.2025: Grok AI by X application on phon

AI/ML

News

Bad actors exploit X’s Grok to run ‘racy’ malvertising campaign

Artificial Intelligence Unleashed: Human Brain on a Technological Canvas, Cyber Space Connection

Application security

AI benefits/risks

OWASP

OWASP warns of ‘unbounded consumption’ risks in AI models

<body><h3>Interview with Harish Peri from Okta</h3>

Oktane Preview: building frameworks to secure our Agentic AI future

Like it or not, Agentic AI and protocols like MCP and A2A are getting pushed as the glue to take business process automation to the next level. Giving agents the power and access they need to accomplish these lofty goals is going to be challenging, from a security perspective. 

How do put AI agents in the position to perform broad tasks autonomously without granting them all the privileges? How do we avoid making AI agents a gold mine for attackers - the first place they stop once they hack into our companies? These are some examples of the questions Okta aims to answer at this year’s Oktane event, and we aim to kick off the conversations a little early - with this interview!

Segment Resources:

<ul>
<li>Check out securityweekly.com/oktane for all our live coverage during the event this year!</li>
<li>More information about the event and how you can attend can be found here: <a rel="noreferrer noopener" target="_blank" href="https://www.okta.com/oktane/">https://www.okta.com/oktane/</a></li>
<li><a rel="noreferrer noopener" target="_blank" href="https://www.okta.com/newsroom/articles/ai-at-work-2025--securing-the-ai-powered-workforce/">AI at Work 2025: Securing the AI-powered workforce</a></li>
<li><a rel="noopener" target="_blank" href="https://www.scworld.com/resource/keeping-ai-under-control-what-to-expect-at-oktane-2025">Keeping AI Under Control: What to Expect at Oktane 2025</a></li>
</ul>

<h3>Topic - Indirect Prompt Injection Getting Out of Hand</h3>

Reports of indirect prompt injection issues have been around for a while. Of particular note was Michael Bargury's <a rel="noreferrer noopener" target="_blank" href="https://i.blackhat.com/BH-US-24/Presentations/US24-MichaelBargury-LivingoffMicrosofCopilot.pdf">Living off Microsoft Copilot</a> presentation from Black Hat USA 2024. Simply sending an email to a Copilot user could make bad stuff happen.

Now, at Black Hat 2025, we've got more: the ability to plunder any data resource connected to ChatGPT (they call these integrations "Connectors") from Tamir Ishay Sharbat at Zenity Labs. The research is titled <a rel="noreferrer noopener" target="_blank" href="https://labs.zenity.io/p/agentflayer-chatgpt-connectors-0click-attack-5b41">AgentFlayer: ChatGPT Connectors 0click Attack</a>.

Looks like Google Jules is also vulnerable to what the Embrace the Red blog is calling <a rel="noreferrer noopener" target="_blank" href="https://embracethered.com/blog/posts/2025/amp-code-fixed-invisible-prompt-injection/">invisible prompts</a>. Sourcegraph's Amp Code <a rel="noreferrer noopener" target="_blank" href="https://embracethered.com/blog/posts/2025/amp-code-fixed-invisible-prompt-injection/">is also vulnerable to the same attack</a>, which encodes instructions to make them invisible.

What's really going to ruffle feathers is the fact that all these companies know this stuff is possible, but don't seem to be able to figure out how to prevent it. Ideally, we'd want to be able to distinguish between intended instruction and instructions injected via attachments or some other means outside of the prompt box. I guess that's easier said than done?

<h3>News</h3>

Finally, in the enterprise security news, 

<ol>
<li>Drones are coming for you… to help?</li>
<li>One of the most powerful botnets ever goes down</li>
<li>Phishing training is still pointless</li>
<li>Microsoft sets an alarm on its phone for 8 years from now to do post-quantum stuff</li>
<li>vulns galore in commercial ZTNA apps </li>
<li>GenAI projects are struggling to make it to production</li>
<li>Adblockers could be made illegal - in Germany</li>
<li>Windows is getting native Agentic support</li>
<li>Automating bug discovery AND remediation?</li>
<li>Public service announcement: time is running out for Windows 10</li>
</ol>

All that and more, on this episode of Enterprise Security Weekly.

This segment is sponsored by Oktane by Okta. Visit <a rel="noreferrer noopener" target="_blank" href="https://securityweekly.com/oktane">https://securityweekly.com/oktane</a> to learn more about them!</body>

Featured image for {“vendor”:”ppworks”,”type”:”segment”,”id”:”14164″} podcast from PPWorks

Identity

Oktane

Oktane Preview with Harish Peri, Invisible Prompt Attacks, and the weekly news! – Harish Peri – ESW #421

gpt5_AdobeStock_1635680951_Editorial_Use_Only

Exposure management

GPT-5 jailbreaks reported despite OpenAI’s new safety-training method

large language models Concept illustration Ai

Feature

LLM vector flaws threaten data security, privacy, and model integrity

Four hackers gather around a computer screen in a still from the movie "Hackers."

hackers-still-mgm

Black Hat

Risk Identification/Classification/Mitigation

Black Hat: Sloppy AI defenses take cybersecurity back to the 1990s, researchers say

<body>Maintaining code is a lot more than keeping dependencies up to date. It involved everything from keeping old code running to changing frameworks to even changing implementation languages. Jonathan Schneider talks about the engineering considerations of refactoring and rewriting code, why code maintenance is important to appsec, and how to build confidence that adding automation to a migration results in code that has the same workflows as before.

Resources

<ul>
<li><a rel="noreferrer noopener" target="_blank" href="https://docs.openrewrite.org">https://docs.openrewrite.org</a></li>
<li><a rel="noreferrer noopener" target="_blank" href="https://github.com/openrewrite">https://github.com/openrewrite</a></li>
</ul>

Then, instead of our usual news segment, we do a deep dive on some recent vulns NVIDIA's Triton Inference Server disclosed by Trail of Bits' Will Vandevanter. Will talks about the thought process and tools that go into identify potential vulns, the analysis in determining whether they're exploitable, and the disclosure process with vendors. He makes the important point that even if something doesn't turn out to be a vuln, there's still benefit to the learning process and gaining experience in seeing the different ways that devs design software. Of course, it's also more fun when you find an exploitable vuln -- which Will did here!

Resources

<ul>
<li><a rel="noreferrer noopener" target="_blank" href="https://nvidia.custhelp.com/app/answers/detail/a_id/5687">https://nvidia.custhelp.com/app/answers/detail/a_id/5687</a></li>
<li><a rel="noreferrer noopener" target="_blank" href="https://github.com/triton-inference-server/server">https://github.com/triton-inference-server/server</a></li>
<li><a rel="noreferrer noopener" target="_blank" href="https://blog.trailofbits.com/2025/07/31/hijacking-multi-agent-systems-in-your-pajamas/">https://blog.trailofbits.com/2025/07/31/hijacking-multi-agent-systems-in-your-pajamas/</a></li>
<li><a rel="noreferrer noopener" target="_blank" href="https://blog.trailofbits.com/2025/07/28/we-built-the-security-layer-mcp-always-needed/">https://blog.trailofbits.com/2025/07/28/we-built-the-security-layer-mcp-always-needed/</a></li>
</ul></body>

Featured image for {“vendor”:”ppworks”,”type”:”segment”,”id”:”14098″} podcast from PPWorks

DevOps

Uniting software development and application security – Jonathan Schneider, Will Vandevanter – ASW #342

ai_AdobeStock_1364146916

Data Security

Almost a quarter of the files workers upload to AI has sensitive info