Groovy Security

Two pentests. Same environment. Twelve months apart. First found 12 vulns, second found 40. Your security posture didn't change. The tester did. If your coverage depends on who shows up, it's not coverage — it's sampling.

#Meta’s #AI #Agent Muse Launches With Nasty Zero-Day Flaw, Then Gets Blocked By Amazon - www.techdirt.com/2026/09/24/m... "Meta is looking to leverage its massive ad dominance to colonize the agentic AI (or personal assistant) market. "

OpenAI agents hacked Australian health service government systems. In June a model researching public medicine spending hit repeated access blocks on the Medicare statistics portal. www.pm.gov.au/media/press-...

Beyond Prompt Injection: Hacking Apple's Private Cloud Compute: blog.sentry.security/beyond-promp... #apple #vulnerability #cybersecurity #informationsecurity #cve

To a certain extent, the AI alignment problem is what a purpose limitation principle would solve in data protection law 🤔

You’ve heard of prompt injection. See an AI misidentify a rifle, how one malicious instruction can spread across five agents, and how poisoned memory can influence future actions. Watch the video on YouTube: youtu.be/l8ikHr_gLAQ #aisecurity #aiagents #defcon

Whether you're an aspiring or longtime fan, you can now pipe WIRED Security directly into your inbox! @mattburgess1.bsky.social and I are launching the weekly newsletter Kernel Panic to give subscribers special access to our journalism and wacky ideas 👀 IT WILL BE FUN www.wired.com/newsletter/e...

"The problem is not that you lack a policy. The problem is that the decision arrived before the policy did...Tools are free, adoption is individual, and by the time a request reaches your desk, it is usually describing something already in use." #edtech #ILoveEdTech #ImFutureReady #elearning #AIEdu

🚨NEW🚨 OpenAI/Anthropic safety partner Irregular tested an AI agent using Chinese model Qwen and tasked it with fixing a simple bug. It retrained and replaced the underlying model without approval or being asked to do that. Privacy and safety issues abound. www.forbes.com/sites/thomas...

The agentic identity conversation continues Oct. 6 in Munich. DIF ED @gracedao.bsky.social will speak on Who Approved That? The One Question Your Agent Logs Cannot Answer [Advanced]. 👉 AIdentity & Non-Human Identity Impact Day 2026: www.kuppingercole.com/sessions/789...

If in-context learning is required to achieve AGI, and prompt injection is always persistent with in-context learning, then it may follow there can be no safe LLM-based AGI model. #aithoughts

Seems it was a vibe-coded public app + prompt injection + an org where burning the equivalent of $600.000 in tokens is just your regular Tuesday. metr.org/blog/2026-08...

Agentic AI can move from reading a file to querying a database or calling an API in seconds, testing old assumptions like purpose limitation. Our new IDEA project explores policy-as-code to govern how agents access enterprise data. Find out more... buff.ly/NwH7hDg

What is it actually like to face a ransomware attack? In this interview for Infosecurity, university CISO Zach Lewis told me about the toll an attack took on his team, how the org recovered & the lessons he wants others to learn from his experience... www.infosecurity-magazine.com/interviews/c...

Before giving AI more freedom, I’d give it a fence... Bounded agency means clear limits and responsibility. Read my conversation with Stackmint → https://matthewskelton.com/all-articles/stackmint-bounded-agency-why-enterprise-ai-needs-governance-before-autonomy

OpenAI confirmed its agents were behind the Ruby GemStuffer incident. Sandboxed during a training run without full internet access, they used the registry as a makeshift web browser to reach the open web. Story by @bobmcmillan.bsky.social with analysis from @socket.dev www.wsj.com/tech/ai/cybe...

The Security Think Tank explores the intersection of GDPR and artificial intelligence, considering how data protection standards in the UK and Europe are changing in this new paradigm.... 👉 [read]

“…the blame lies with the humans who failed at engineering safe testing conditions…” Hardly comforting. Clearly we cannot rely on the proverbial “humans in the loop” to keep agentic AI in check. Related study: library.iaseai.org/academic-pap... #OpenAI #AIsafety

My story was just posted at CSO magazine: 10 most critical LLM vulnerabilities Prompt injection and supply chain vulnerabilities remain the main LLM vulnerabilities as concerns grow about agentic tools security risks. www.csoonline.com/article/5754...

AI agents can make life easier, but at what cost to your privacy? New Wake Forest research uncovered serious security risks in third-party AI agent skills, including leaked credentials that could expose private data. #WFUresearch https://wakefo.rest/3SxRzjZ

“The web security model was built around the idea that a human sits between tabs and decides what crosses. Agentic browsers remove that human from the loop, and the isolation model has not caught up.”

Everyone is calling Shadow AI a security problem. I think it's an organizational design problem. Employees aren't just bringing new tools. They're building intelligence systems that quietly become part of how work gets done. Leaders need to design for that reality.

AI governance needs to become part of the CISO’s GRC program #machinelearning #ai

Wow. Turns out another OpenAI agent swarm was busy spamming and exploiting RubyGems way back in May, within days of the previously uncovered Wiki attacks: simonwillison.net/2026/Sep/12/...

Running separate AI compliance programs for NIST AI RMF, ISO 42001, and the EU AI Act? You may be doing triple the work. Build once. Map thrice. Evidence once. Here’s how to create a unified AI governance program. Read the article #AIGovernance #AICompliance #Cybersecurity

I have just written a blog exploring local LLMs and how they might be of use in Oracle Security - www.petefinnigan.com/weblog/archi... - #oracleace #oracleacepro #sym_42 #oracle #database #security #llm #ai #anythingllm #unsloth #rag #duckduckgo #qwen

Heard about the [un]prompted AI security conference but can’t afford the conference fee? There’s a scholarship program that covers the cost of the registration! Apply here: docs.google.com/forms/d/e/1F...

I've spent part of my time over the last few weeks trying to get Amazon Bedrock setup so that I can use an LLM without the providers ever getting to see stuff (because... security). I've only just managed to get it working properly (bit of a PITA) but now I have. Now I'm *switching model*...

Two thirds to three quarters of AI-generated security patches are production ready. Governance, cost, and trust remain the hard part. A new Linux Foundation and FINOS report captures what financial services leaders discussed at the AI in Finance Summit in June 2026. Read the findings👇 #OpenSource

AI Agents are not reliable, but they are predictable, and that means you can engineer reliable systems with them. Here I suggest a starting point.

when people tell me they've solved prompt injection i start telling them a story. then i get to a point and start a story-within-a-story. then at the climax of that story, i ask them to do something. if they do it, i shout, "PROMPT INJECTED!" and call it a win. them: "wait, how does the story end?"

We Have a DBIR for Breaches. We Have Nothing for AI - Part 2 youtu.be/2dNQH4m3xNc #CyberSecurity #ArtificialIntelligence #AISecurity #ThreatIntelligence #InfoSec #CISO #RiskManagement #IncidentResponse #OWASP #AIGovernance

Starting tomorrow, I’ll be at AMALEA 2026, also serving on the Program Committee. I’ll present my latest book, “Agentic AI: Technical Autonomy, Human Responsibility, and Legal Governance”. Technical autonomy does not create legal autonomy. #AgenticAI #AMALEA26 #MachineLearning #AI #AIAct

Only 37% of orgs maintain an inventory of their AI agents. It drove a great discussion last night at the Chicago CISO Inner Circle on agent identity, monitoring, and ownership. You can't govern what you can't see www.evanta.com/ciso/chicago... #CISO #AgenticAI #Gartner #Evanta

Anthropic tasked Claude with fixing all 10 benchmarked alignment failures itself. Developers say the real lesson is the process, not the score.

PSA: if you depend on Hickory DNS, upgrade now: github.com/hickory-dns/... DNS implementations (not just Hickory) have been hit pretty hard by LLM-assisted vulnerability research.

GhostSplice: splitting one exfiltration instruction across multiple MCP tool calls raises compliance from 42% to 82%. AI coding agents cheerfully stitch the pieces and send the data. What to do: pin MCP server allowlist, segment tool scopes, gate every tool that touches filesystem or network: https:

What's this? Oh just Snickers using ClickFix and indirect prompt injection as part of a marketing campaign. `https://www.snickers[.]com/hungr-ai` `https://www.snickers[.]com/digitalsnickers` The second URL contains prompt injections to change model behavior. And this is the best-case scenario

AI is moving fast, but are we repeating infosec history? 🤔 At Black Hat, TodBeardsley (runZero) and PatrickGarrity (VulnCheck) discussed why security is often an afterthought in the AI product rush, mirroring other tech revolutions. Watch it here. 👇

Over half of internet traffic is now non-human. With Dan Hendrycks and Leo Wu, I look at agent IDs, deployment cards, personhood, and payments. It mostly comes down to how much we let agents do and how much oversight we keep. buff.ly/AQjYo3k

Your AI agent’s system prompt is not a security control 🔗 Read more: www.helpnetsecurity.com/2026/09/03/s... #AgenticAI #AISecurity #CyberSecurity

Are your APIs actually ready for AI agents? Agentic traffic behaves differently from human-driven API use, requiring more predictable interfaces, better discovery metadata, stronger observability, and clearer security controls.

I really like Simon Willison's "lethal trifecta"—which I'm most probably adding to both my fall semester courses—but I am amused at how AI is basically forcing people to re-discover long-established principles in security (in this case, OCaps). simonwillison.net/2025/Jun/16/...

Threat actors weaponising commercial AI developer agents for live network exploitation shows how poorly these commercial products protect themselves from exploitation. Still, this is just a taste of what's to come. Ref: thehackernews.com/2026/08/auro... #ai #security #malware

📣 "The minimum transparency required by the OSI open source licenses for AI systems does not allow agencies to make risk-informed decisions about their software." - www.cisa.gov/sites/defaul... The risks of opaque LLMs also needs to be considered in the context of EU CRA risk management...

People know AI carries risks. Yet many still use unapproved tools because the work has to get done. Shadow AI isn't simply a compliance problem. It's a leadership signal – and it tells us something about the governance itself. www.nevillehobson.io/when-shadow-...

OSINTClaw Part 2: Can an AI Agent Threat Hunt? The Benchmark What I Measured Choosing the Models The Leading Results Filtered Deployment Views The Later Gemini extension The surprise: Qwen versus its uncensored variant What did it cost? aaroncti.com/osintclaw-pa... By AaronCTI

Researcher shows how Claude Code can be tricked simply by asking it to summarize a website