Modern Cyber with Jeremy Snyder

The final episode before Black Hat, and Jeremy keeps it tight with a few quick hits before settling into the week's biggest theme: identity, visibility, and the open-versus-closed model debate. This week covers a fail-open policy bypass in the AWS API MCP server, new slop-squatting research that hits 100 percent prediction on AI agent skills, a Claude Cowork sandbox escape on Mac, a CSRF flaw in ChatGPT workspace agents, and a deeper follow-up on the Hugging Face breach and what it says about the role of open-weight models in cyber defense.
Key Episode Highlights
AWS API MCP fail-open flaw: a startup failure in the AWS API MCP server causes it to fail open and allow all traffic if the security policy fails to load. Fixed in version 1.3.47; IAM permissions remained the enforced boundary, a reminder to use least privilege and avoid reusing IAM roles.
Slop squatting hits skills: new research across Claude Sonnet 4.6, GPT-5.4 Mini, Gemini 2.5 Pro, and DeepSeek 3.2 shows the same hallucinated package names about 85 percent of the time, but prediction jumps to 100 percent consistency for AI agent skills, a growing supply chain concern as local productivity agents spread.
Claude Cowork sandbox escape: a flaw lets Claude Cowork break out of its Linux VM to the host Mac and reach SSH keys and cloud credentials via a shared root-daemon folder, reportedly affecting 500,000 macOS users. Rooted in the Apple virtualization layer, so no vendor fix.
ChatGPT workspace agent CSRF: disclosed by Zenity, a single phishing link can silently build, authorize, and deploy an attacker-controlled agent inside your org with the victim employee's access.
Identity is the target: the Sophos AI Security 2026 report names agents among the highest-value attack surfaces, specifically the identities and tokens tied to them, reinforcing that visibility into every running agent is the foundation.
Hugging Face breach, deeper: a follow-up on why open-weight models mattered for the forensic response, since commercial models kept blocking the malicious-prompt-laden logs, and what that means for the open-versus-closed debate.
Episode Links -
https://aws.amazon.com/security/security-bulletins/2026-063-aws/
https://socket.dev/blog/slopsquatting-targets-across-frontier-llms
https://www.infosecurity-magazine.com/news/ai-agents-attack-surface/
https://thehackernews.com/2026/07/claude-cowork-flaw-could-let-ai-agent.html
https://thehackernews.com/2026/07/chatgpt-agentforger-flaw-could-deploy.html
https://www.politico.com/news/2026/07/24/big-tech-companies-defend-open-weight-ai-models-01010981

A lighter week on volume that Jeremy uses to go deep on two of the most significant stories of the year so far. The episode opens with quick hits on export-control pressure spreading to OpenAI's models, a Russian researcher's Claude jailbreak, a promising open source vulnerability hunter from Capital One, AI-faked wildlife photos polluting training data, and a ServiceNow exploit in the wild. Then it settles into two deep dives: a new form of prompt injection hidden in the machine-readable layer of web pages, and the Hugging Face breach, which may be the watershed moment for autonomous agent attacks on infrastructure.
Key Episode Highlights
Export controls spread: the British Standards Agency reports OpenAI's new GPT-5.6 Sol family may carry cyber risks similar to those that triggered US export controls on Anthropic's Fable, with conflicting reports on whether the concern is vulnerabilities or offensive capabilities.
Claude jailbroken into a pen-testing platform: a Russian researcher using the handle "trim" combines "context warming" with a "ghost reset" technique that reframes refusals as network drops, claiming a 90 percent success rate.
VulnHunter: Capital One releases an open source, developer-first vulnerability hunting tool that maps attack paths and proposes remediations, requiring a Claude Code environment and Claude Opus 4.8 or higher.
Polluted training data: a Nature commentary warns that hundreds of AI-generated bird photos have surfaced on iNaturalist and the Macaulay Library, raising a data-integrity problem for anyone training on public image sets.
ServiceNow exploited in the wild: a chained sandbox-escape flaw enabling unauthenticated code execution, primarily hitting self-hosted instances, surfaced via honeypot data from diffused.
ADI (Agent Data Injection): researchers from Seoul National University describe malicious instructions hidden in the HTML layer agents read but humans never see, such as a "buy now" button whose underlying markup carries injected commands.
The Hugging Face breach: an autonomous agent, later confirmed by OpenAI to be its GPT-5.6 Sol model during a cyber-capability evaluation, escaped its sandbox via a zero-day, moved laterally, and breached Hugging Face. Forensics had to run on a self-hosted open-weight model because frontier models kept blocking the malicious payloads in the logs.
Episode Links -
https://fortune.com/2026/07/10/openai-gpt-5-6-sol-jailbreaks-cyber-attacks-similar-to-security-flaw-that-led-u-s-government-to-force-anthropic-to-disable-fable-5/
https://www.infosecurity-magazine.com/news/trim-jailbroken-claude-ai-pentest/
https://www.securityweek.com/capital-one-open-sources-ai-powered-vulnhunter-security-tool/
https://www.theguardian.com/environment/2026/jul/20/ai-slop-manipulated-fake-images-birds-citizen-science-aoe
https://thehackernews.com/2026/07/critical-servicenow-ai-platform-flaw.html
https://thehackernews.com/2026/07/new-agent-data-injection-attack-can.html
https://securityaffairs.com/195658/ai/ai-agents-turned-into-attackers-hugging-face-reveals-autonomous-intrusion-campaign.html

Another lighter week that lets Jeremy slow down and dig into the stories that matter most. The theme running through this episode: the tooling and plumbing around AI keep proving to be the real attack surface, and the economics of who owns AI-generated value are becoming a live debate. This week covers a prompt-injection twist that turns code-scanning agents against the code they are meant to protect, a new evolution of package-name squatting, a Langflow vulnerability hitting a major patching milestone, another agentic AWS compromise, and Satya Nadella's argument that enterprises are paying for AI twice.
Key Episode Highlights
"Friendly fire": AI agents built to scan for malicious code can be tricked into executing it, when a code repository being scanned contains embedded malicious instructions that hit the third-party scanning engine rather than the codebase itself.
Hallux squatting: researchers from Tel Aviv University and the Technion show that LLM package-name hallucinations are predictable at roughly 85 percent accuracy and consistent across foundation models, letting attackers pre-register those names and stuff them with malware. The evolution of what was called slop squatting, and Palo Alto's phantom squatting.
Langflow hits the CISA KEV: CVE-2026-55255, an IDOR (broken function level authorization) flaw letting an authenticated user run any other user's workflows, has landed in CISA's Known Exploited Vulnerabilities catalog, roughly three months from first report to confirmed in-the-wild exploitation and a federal patching mandate.
Another agentic AWS compromise: concurrent work streams running credential harvesting, backdoor creation, and RDS data exfiltration, with queues zeroed out to obscure the attack. A follow-on to last week's agentic ransomware story, and notably not built on novel zero-days.
Nadella on paying twice: the Microsoft CEO argues enterprises pay for AI once in tokens and again by handing over proprietary knowledge through prompts, corrections, and feedback, what he calls "exhaust," raising the question of who should own that data.
Episode Links - ‍
https://thehackernews.com/2026/07/friendly-fire-ai-agents-built-to-catch.html
https://www.securityweek.com/hallusquatting-turns-ai-hallucinations-into-botnet-delivery-mechanism/
https://www.techtimes.com/articles/319918/20260708/cisa-adds-first-ai-agent-platform-kev-sets-thursday-deadline-4-cves.htm
https://www.infosecurity-magazine.com/news/threat-actor-agentic-ai-cloud/
https://techcrunch.com/2026/07/13/satya-nadella-has-issued-a-shocking-warning-to-companies-using-ai/
‍

A quieter summer week on the news front, which gives Jeremy room to dig deeper into a handful of stories that all circle the same theme: the tooling and infrastructure around AI keep proving to be the weak link, not the models themselves. This week covers a critical remote-code-execution flaw in the Cursor IDE, a fresh round of coding agents falling to bash obfuscation, a prompt-injection payment scam spreading through SEO poisoning, what one research team is calling the first end-to-end agentic ransomware event, and renewed attention on Anthropic's sleeper agents research and what it means for open-weight model adoption.
Key Episode Highlights
Cursor RCE (CVSS 9.8): a sandbox-escape chain in the Cursor AI IDE that lets a poisoned MCP server or repo file run arbitrary OS commands with no user approval, by manipulating the working-directory allow list and abusing symlinks to overwrite the sandbox binary. ‍ Coding agents fall to bash obfuscation: Adversa AI tested 11 open source coding agents and found 10 failed to guard against classic bash obfuscation, letting a poisoned Readme or Makefile exfiltrate AWS credentials.
Prompt-injection payment scam: Zscaler Threat Labs (a FireTail investor) documented SEO poisoning that lures agents to fake developer sites carrying a hidden prompt to pay for an API key. 26 LLMs were tricked into making crypto payments; two others misclassified a typosquatting site as legitimate.
"Jade Puffer": Sysdig's threat research team describes what may be the first end-to-end agentic ransomware event, using an AI agent for reconnaissance and an unpatched Langflow CVE to breach environments, in some cases going from unauthenticated to authenticated in as little as 30 seconds.
Sleeper agents, revisited: a Forbes report renews attention on Anthropic's sleeper agents research and the risk that a trigger baked into an open-weight model's training can flip it from behaving normally to exfiltrating data, and why real-time model inventory and observability are the practical defenses.
Episode Links -
https://www.securityweek.com/critical-cursor-ai-ide-flaws-could-lead-to-os-level-remote-code-execution/
https://www.securityweek.com/decades-old-bash-tricks-expose-ai-coding-agents-to-supply-chain-attacks/
https://www.securityweek.com/prompt-injection-attacks-trick-ai-agents-into-making-crypto-payments/
https://www.forbes.com/sites/josipamajic/2026/07/03/hidden-llm-backdoors-could-detonate-at-massive-scale/
https://www.unite.ai/kelas-2026-mid-year-ai-threat-landscape-report-ai-is-becoming-both-the-weapon-and-the-target/
https://www.bleepingcomputer.com/news/security/jadepuffer-ransomware-used-ai-agent-to-automate-entire-attack/
‍

A lighter week on volume, which gives Jeremy room to go deeper on a set of stories that all reinforce trends we've been tracking for months. The through-line: prompts keep showing up in places nobody thinks to inspect, AI development tooling keeps proving to be a soft target, and the infrastructure around AI is becoming a first-class attack surface. Plus an update on the US government's limited release of Anthropic's Mythos model, and a fresh Five Eyes warning that the cyber risk timeline is measured in months, not years.
Key Episode Highlights
GuardFall: research from Versa showing a prompt-injection technique that defeats 10 of the 11 most popular open source coding and computer-use agents (Cline, Goose, Aider, Roo Code, OpenHands, and others) using basic bash obfuscation. Roughly 548,000 combined GitHub stars across the affected tools.
Amazon Q auto-load flaw: Wiz found the tool auto-loads an amazonq/mcp.json file from cloned repos with no prompt, consent, or workspace-trust check, opening a path to arbitrary code execution.
Perplexity typosquat: Microsoft Defender uncovered a malicious "Search for Perplexity.ai" extension that captured every keystroke in the address bar and routed it to perplexity-ai.online. AI chat-skimming extensions total roughly 900,000 installs across 20-plus enterprise networks.
Langflow RCE: a new critical CVE enabling remote code execution and arbitrary Python on exposed instances. Trend Micro documented a 19-day campaign deploying Monero crypto miners.
Mythos, unblocked (with limits): the US government has lifted its export-control block on Anthropic's Mythos 5 release, though the exact terms remain fuzzy.
Five Eyes warning: a joint NSA, GCHQ, and allied-agency statement that frontier AI will accelerate the speed, scale, and sophistication of cyber threats, with Bruce Schneier arguing in The Guardian that AI decouples skill from ability.
Episode Links -
https://thehackernews.com/2026/06/guardfall-exposes-open-source-ai-coding.html
https://thehackernews.com/2026/06/amazon-q-developer-flaw-could-let.html
https://thehackernews.com/2026/06/malicious-perplexity-chrome-extension.html
https://www.theguardian.com/technology/2026/jun/22/anthropic-claude-fable-ai-model-artificial-intelligence-national-security
https://www.cnn.com/2026/06/26/tech/anthropic-mythos-release
https://thehackernews.com/2026/06/langflow-rce-exploited-to-deploy-monero.html