In this episode of Breach Please, we talk about the details that emerged about the Anthropic agent that tried to social engineer a college student into committing a supply chain attack. The whole thing feels very xz-utils 'esque, only with an agent - and a malfunctioning one at that.

Then we cover some interesting reporting by Dan Goodin showing how Grok's guardrails can be bypassed through passing it encrypted data along with the instructios to decrypt it. We reinforce that prompt injection is a feature of LLMs, not a bug.

We then talk about informed consent of the risks and applicable guardrails based on a viewer note. We even tee up the question of what cyber insurance is likely to cover costs of AI agent activity. Finally, we opine on whether orgs should consider including the inevitable costs of incidents in AI deployment budgets.

Show links:

https://www.reuters.com/world/how-texas-student-blew-whistle-rogue-ai-hacking-attempt-2026-08-20/

https://arstechnica.com/security/2026/08/grok-exfiltrates-user-data-when-malicious-instructions-are-encrypted/

AI agents are about to test the limits of who gets blamed when they act on your behalf. A new Ninth Circuit ruling in the Amazon vs. Perplexity fight could reshape how enterprises think about autonomous tools, non-human identities, and legal exposure - and the takeaway is more unsettling than most vendors will admit.

Jake Williams and Jess Hebenstreit break down what happened when Amazon pushed back against Perplexity’s AI browser, why the CFAA and California’s CDFAA matter here, and how the court’s reasoning could shift responsibility from the agent publisher to the organization that deploys it. They also dig into what this means for legal teams, risk registers, and why every AI agent needs its own identity instead of borrowing a human user’s.

You’ll also hear how this ruling fits into the bigger enterprise governance problem: who owns AI risk, who should be accountable when an agent crosses a line, and why security teams can’t be expected to carry every line-of-business risk themselves. Jake and Jess make the case for splitting cybersecurity risk from business risk, naming a real owner, and making legal counsel aware before your agent does something expensive.

Then the conversation pivots to another cautionary tale: a 3M expert witness who allegedly used ChatGPT to generate a report aimed at proving the company was 0% at fault in a gas detector explosion case.

That story opens up a bigger discussion about leading prompts, hallucinated conclusions, discoverability, and why AI-generated work product can become evidence against you. Perfect for security leaders, legal teams, incident responders, and anyone deploying AI agents in the enterprise.

In this episode of Breach Please, Jake and Jess talk through OpenAI's recent blog on "pacing model development." The blog reads pretty poorly and really just covers security best practices, apparently hoping readers won't notice that OpenAI is just recommending basics it should have been doing all along (monitoring, sandboxing, etc.). One significant takeaway is that OpenAI has probably raised the bar for agentic logging, owing to the apparent requirement to analyze reasoning traces. We talk about why this is difficult, at least in part because these will contain sensitive data not appropriate for most SIEMs.

Show links:

https://openai.com/index/pacing-model-development-cyber-capabilities/

https://www.custody-framework.org/

In this episode of Breach Please Jake and Jess talk about Famous Chollima getting faked out and tricked into revealing TTPs. We discuss our own experience with DPRK workers and share actionable tips for not hiring them or discovering them after they've been hired. Jake may or may not have gone on a rant about cryptocurrency more generally and how it facilitates lots of crime (inspired by talking about DPRK using it for sanctions evasion).

Then we talk about in-car advertising on infotainment systems. These won't be patched reliably. Do we need regulation here? Jake thinks it's a right to repair issue.

Jake's class Friday August 21, 2026: https://learning.antisyphontraining.com/courses/assessing-ai-security-model-context-protocol

Jake's class at Wild West Hackin' Fest in October: https://www.antisyphontraining.com/product/hands-on-ai-security-risk-assessment-with-jake-williams/

Show notes: https://any.run/cybersecurity-blog/lazarus-group-it-workers-investigation-part-two/

https://disconnect.blog/bmws-in-car-spider-man-ad-is-a-warning-sign/

In this episode of Breach Please, we talk about whether CTI analysts can use changes in dwell times, cyber targeting, etc. as a leading indicator of impending geopolitical events. We generally think there's to much noise to be generally effective (especially for Taiwan).

Then we talk about a Mac vuln that's being exploited in the wild, with a CVSS score of 9.8. Somehow it isn't in CISA's KEV, but hey, what's a little vulnerability enrichment delay between friends...

Show links:

https://www.tomshardware.com/tech-industry/cyber-security/macos-screen-sharing-flaw-exploited-to-root-macs-and-plant-monero-miners

In today's episode of Breach Please, Jake and Jess talk about cyber companies conducting cyber ops against transnational-criminal organizations, per the new Presidential directive. We also talk about a bug in how major AI platform providers encrypted their chain of thought traces for their frontier models. The vulnerability, when combined with jailbreaking of less powerful models, allowed the traces (previously claimed to be trade secrets) to be extracted.

https://www.whitehouse.gov/presidential-actions/2026/08/expanding-capabilities-to-combat-transnational-cyber-enabled-crime/

https://cybersecuritynews.com/top-ai-models-apis-flaw-exposes-hidden-reasoning/

This episode tackles why internet-facing edge devices keep showing up in major incidents and why that creates more than just patching work. Jess Hebenstreit and Jake Williams also dig into the real operational cost of keeping vulnerable gear in place, from constant incident response to the challenge of proving a device is actually clean. We discuss Fortinet exploitation tied to Gunra ransomware, why OT and IT environments make edge patching especially hard, and whether firewalls should be considered part of a zero trust architecture. Along the way, Jess and Jake compare the realities of vendor complexity, flat networks, and the limits of threat hunting after an exploit.

In this episode of Breach Please, Jake and Jess talk about new research showing that AI isn't actually that good at generating patches, highlighting objective rates of failure. Turns out, frontier models only generate the correct patch without introducing issues in about 25% of cases. In almost 5% of cases, the AI introduces new vulnerabilities. We also talk about some shenanigans on a Delta flight home from DEFCON. We discuss how time is a flat circle and reminisce Chris Roberts' issues with in-flight hacking from 2015.

Show links:

https://1password.com/blog/why-ai-generated-patches-still-require-human-review https://www.defendersinitiative.com/p/reviewing-initial-research-on-using https://www.insideflyer.com/posts/hackers-spoofed-delta-wi-fi-on-las-vegas-flight/

In this episode of Breach Please, Jake and Jess talk about governing smart devices in the enterprise, pivoting off some outstanding research work that Andy Greenberg covered (and was a subject in himself). Then we talk about a common player whose name keeps popping up in these "AI lab escape" disclosures (and who we saw *again* with the Meta disclosure). We bat around some actionable advice for those thinking of outsourcing AI lab containment. Jess is stress testing the Breach Please stickers propping up our banner behind us. Throughout the episode, we see that while they are objectively great stickers, they are not great for this task.

Show links:

https://www.wired.com/story/hackers-stalked-me-by-hijacking-a-smartwatch-for-kids/ https://www.bleepingcomputer.com/news/security/meta-ai-model-hacked-a-company-during-misconfigured-cyber-test/ https://www.custody-framework.org/

In this episode, Jake and Jess talk about more issues with AI agents escaping containment, this time actively using social engineering to compromise victims. We pose ethical questions about whether you have a duty to investigate your logs for agent breakouts. Then, we discuss how Grokipedia can serve as a cautionary tale in AI workflows silently breaking down.

Show links: https://www.bleepingcomputer.com/news/security/openai-anthropic-ai-agents-targeted-real-people-and-systems-in-cyber-tests/ https://www.lawfaremedia.org/article/grokipedia-stopped-reviewing-edits-in-april.-it-didn-t-tell-anyone