One Image to Root Them All - The 443 Podcast - Episode 388

This week on the podcast, we break down how a heap overflow in an image-decoding library nobody thinks about became a pull request inside OpenAI's internal monorepo. Three researchers chained it with an OpenAI single sign-on flaw to hijack employee ChatGPT and Codex accounts in under 72 hours; and had an AI write the exploit for them. Before that, we cover the actively exploited maximum-severity authentication bypass in Cisco's Identity Services Engine. Then we close with RatHat, a new Android malware that enables Wireless Debugging, reads its own pairing code from the screen, and asks a commercial AI assistant where to tap.

This week on the podcast, we analyze a set of actively exploited vulnerabilities in Mikrotik's RouteOS followed by a pair of actively exploited vulnerabilities in Cisco's Firewall Management Center. After that, we review a phishing attack against Trezor hardware cryptocurrency wallet users that originated with their third party marketing email provider.

This week, dive into OpenAI's and METR's analysis of the rogue OpenAI agents that hacked Hugging Face back in July. We go over the full attack timeline discussing the multiple message boards the agents created and the ultimate goal of their attack against Hugging Face and trust us, its as crazy as it gets. Before that, we discuss a recent dark web service that popped up selling 153 million stolen drivers licenses before discussing a research post into a malware implant discovered in inexpensive Chinese routers.

This week we sit down with Euler Neto, a cybersecurity analyst at WatchGuard, to discuss his research into the ErrTraffic Malware-as-a-Service loader found targeting Portoguese-speaking users in recent months.

This week on the podcast, we discuss a research post that defines a new attack technique against AI tools called Cryptographic Context Injection. Before that, we cover an authentication bypass vulnerability in CircleCI's MCP server after discussing 3.6 million records stolen from the Azure tenants of several large organziations.

This week on the podcast, we discuss a new White House memorandum that creates a program to authorize American private companies to begin conducting offensive cyber operations. Before that, we discuss a vulnerability write up for a Citrix Netscaler flaw before covering yet another prompt injection vulnerability in a popular AI tool.

This week on the podcast, we cover a series of alerts form US law enforcement and intelligence sources describing Iran-backed attacks against municipal water supply utilities. Before that, we give an update on the OpenAI and HuggingFace rogue AI saga before discussing a research post on MCP configuration file risks.

This week on the podcast, we review HuggingFace's technical write up of their recent run in with a rogue OpenAI model, as well as their CEO's demands from OpenAI in response. We then cover an interesting research whitepaper that describes a side channel attack that could let AI transcribe typed text by an audio recording alone. We end with a threat intelligence report about DNS Poisoning attacks against hotel Wi-Fi systems.

Enter our AI Innovation Challenge: https://www.research.net/r/WatchGuard-AI-Innovation-Challenge-2026   This week on the podcast, we cover the crazy saga that unfolded between the popular open-source AI platform Hugging Face and the frontier AI lab OpenAI. After that, we discuss a recent WordPress remote code execution vulnerability WP2Shell and the research process that Searchlight Cyber followed to uncover it sing artificial intelligence. Finally, we end with a quick analysis of Palo Alto Global Protec's authentication bypass vulnerability CVE-2026-0257

Enter our AI Innovation Challenge: https://secure.watchguard.com/aichall...   This week on the podcast we cover the key takeaways from the just-released Cyber Hygiene Report from WatchGuard. After that, we discuss a recent alert from CISA and other international security agencies on state-sponsored attacks against network equipment. We end with an interesting research post on exfiltrating data from Claude's memory.