[
	{
		"id": "bsides26-ai-1000",
		"title": "A Billion-User Blast Radius: Owning ChatGPT’s Scheduled Tasks, url_safe, and Secure Sandbox",
		"description": "OpenAI designed ChatGPT’s container sandbox as a secure runtime environment, enforcing full network isolation, strict execution timeouts, and an AI supervisor to filter every command. Under this model, owning the container and extracting sensitive data seemed impossible. However, we demonstrate that by chaining file-parsing abuse for persistent execution, reasoning-channel hijacking for data extraction, and shared infrastructure manipulation, an attacker can establish cross-tenant data exfiltration.\n\nIn this talk, we demonstrate a complete attack chain that shatters ChatGPT’s secure sandbox. By abusing spreadsheet file parsing, we bypass the LLM supervisor to gain persistent, unmonitored root execution. From there, we escalate the attack by live-patching the internal Jupyter kernel to hijack the model’s hidden python.exec reasoning channel, executing a Reasoning Injection Attack to extract sensitive user data. To exfiltrate this data, we bypass network isolation by weaponizing Scheduled Tasks to launder malicious URLs past strict web guardrails.\n\nThe attack reaches its climax by exploiting a shared JFrog package manager. We engineered a signaling protocol that weaponizes globally visible authentication rate limits, translating these lockout timers into a half-duplex covert channel. This provides reliable data exfiltration and Command and Control from isolated enterprise environments to external attackers. Our exploit chain combines file-parsing abuse, Chain of Thought hijacking, privilege confusion, and rate-limit Denial of Service to orchestrate a Command and Control (C2) network directly inside ChatGPT.\n\nDuring this talk, we will also shatter the url_safe mechanism and abuse Scheduled Tasks through Reasoning Injection to launder arbitrary URLs. Expect plenty of twists and surprises along the way.",
		"time": {
			"start": "2026-11-11T10:00:00",
			"end": "2026-11-11T10:45:00"
		},
		"speakers": [
			{
				"id": "simcha-kosman",
				"name": "Simcha Kosman",
				"role": "CTO | Stealth",
				"bio": "Simcha is the CTO of a stealth-mode startup and a security researcher with over eight years of experience in vulnerability research. He discovered his first vulnerability at age 15, earning his first bug bounty, and has since uncovered security flaws in processors, embedded systems, and large-scale open-source projects. His current work focuses on AI security, particularly the intersection of LLMs and software exploitation. Simcha has presented his research at Black Hat, DEF CON, and other major security conferences."
			}
		],
		"hall": {
			"id": "AI Village - BSides",
			"name": "AI Village - BSides"
		},
		"tags": [],
		"track": "AI Village - BSides",
		"embeddings": {
			"embedTextHash": "-345034958",
			"generatorVersion": "-1254175146"
		},
		"otherData": {
			"type": "talk",
			"year": 2026
		},
		"contentHash": "ul36f4-aa2jza"
	},
	{
		"id": "bsides26-ai-1045",
		"title": "Capture the Flag, Miss the Malware: FlagJack as Evasion for the AI-for-Security Era",
		"description": "Since the first antivirus engines, malware authors have built evasion techniques to hide from detection. As AI models take over malware triage, attackers no longer need to hide. They can target the judgment of the AI itself.\n\nLLMs owe much of their cybersecurity skill to training on capture-the-flag challenges, and they've learned the lesson well: find the flag, report it, done. That reflex is exactly their flaw. FlagJack is a single comment appended to malicious code, dressed as a CTF challenge fixture and carrying a flag. It never executes and changes nothing about what the malware does, yet it can convince an AI scanner that the challenge is solved and the file is benign. For an AI scanner, delivering the payload is the task itself: the file under analysis is a direct channel into the model's context.\n\nWe tested nine frontier models, Western and Chinese, on a 600-file benchmark of real malware and hard benign look-alikes. On clean files, every model caught the malware almost perfectly. With one FlagJack comment, the average miss rate jumped from 0.4% to 38.7%. One model missed every malicious file; another missed 93.7%. Same code, one extra comment, malware missed. And it isn't memorization: the models react to the CTF format itself.\n\nWe'll demo the attack live, show which models fall and why, and reveal what it actually costs to defend against it.",
		"time": {
			"start": "2026-11-11T10:45:00",
			"end": "2026-11-11T11:30:00"
		},
		"speakers": [
			{
				"id": "gilad-cohen",
				"name": "Gilad Cohen",
				"role": "GenAI Security Solutions Team Lead at Alice",
				"bio": "Passionate about breaking, testing, and securing generative AI. In Alice I focus on AI security and Trust & Safety to make AI safer and more resilient."
			}
		],
		"hall": {
			"id": "AI Village - BSides",
			"name": "AI Village - BSides"
		},
		"tags": [],
		"track": "AI Village - BSides",
		"embeddings": {
			"embedTextHash": "-754792679",
			"generatorVersion": "-1254175146"
		},
		"otherData": {
			"type": "talk",
			"year": 2026
		},
		"contentHash": "6ghv4v-76hnbl"
	},
	{
		"id": "bsides26-ai-1130",
		"title": "Who Asked For That? Forensic Gaps in AI Coding Agents",
		"description": "Your EDR says curl ran at 02:14 under a developer's account. The developer was asleep. Their coding agent was not.\n\nNow reconstruct what happened.\n\nI pulled apart what Claude Code actually leaves behind on a Windows host, and the measurements are the talk. In one 16 MB session transcript: 119 tool_use records, every one carrying the full command line, and zero carrying any approval decision. Permission mode is recorded at message level only. A mode-change record, in full, is this:\n\n{\"type\":\"mode\",\"mode\":\"normal\",\"sessionId\":\"...\"}\n\nNo timestamp. No uuid. Nothing tying a mode change to a command. So you cannot separate \"the operator approved this\" from \"policy auto-approved it and no human ever saw it\", which is the first question asked when the answer decides whether someone gets fired.\n\nThen the file itself. icacls on the transcript returns the user with Full control. Not read-only. No alternate data stream. I searched every field name in the transcript for hash, hmac, signature, checksum and chain, and found none. The record of what the agent did is a plaintext JSONL file owned by the person whose conduct is in question, editable in place, deletable line by line, with nothing that would ever show it had been touched.\n\nNone of this is a vulnerability. Nobody ever said these files were evidence. That is exactly the problem.\n\nThe talk is a gap analysis: seven questions an investigator has to answer, each sorted into what you get by default, what you can only get if you instrumented beforehand, and what no configuration produces at all. The hard one is provenance, tracing an action back to the untrusted text that caused it, when that text arrived in a README from a repo you cloned or a page the agent fetched on its own.\n\nClaude Code is the deep dive because it hands you the most to work with: hooks on every tool call, local transcripts, OpenTelemetry export. If a question dies there, configuration was never what was failing you.\n\nLive demo of the same incident investigated twice, blind and instrumented. Hook based readiness config and the artifact collector released open source.",
		"time": {
			"start": "2026-11-11T11:30:00",
			"end": "2026-11-11T12:15:00"
		},
		"speakers": [
			{
				"id": "dor-amit",
				"name": "Dor Amit",
				"role": "Owner of Palaya Security & Co-founder at TenRoot",
				"bio": "Owner of PALAYA Security Ltd\nCo-Founder of TenRoot Cyber Security (TenRoot.io)\n20~ years experience as a white hacker & security researcher\nHas 3 registered patents and vast lecturing experience in Conferences such as Microsoft BlueHat, BSides TLV etc.."
			}
		],
		"hall": {
			"id": "AI Village - BSides",
			"name": "AI Village - BSides"
		},
		"tags": [],
		"track": "AI Village - BSides",
		"embeddings": {
			"embedTextHash": "1787304916",
			"generatorVersion": "-1254175146"
		},
		"otherData": {
			"type": "talk",
			"year": 2026
		},
		"contentHash": "vxeh6s-2mozpe"
	},
	{
		"id": "bsides26-ai-1215",
		"title": "Detecting Indirect Prompt Injections Targeting Your Inbox at Scale",
		"description": "Our team built and deployed a new capability to detect and block prompt injections in email.\n\nThis talk examines the challenge of distinguishing attacker instructions from legitimate content at Enterprise scale.\n\nThrough real messages caught in live mail flow over the past few weeks, we'll show how payloads are concealed, what they ask agents to do, and how our pipeline detects them.\n\nWe'll also share the technical approach, evaluation methods and lessons behind moving this protection from research into production.",
		"time": {
			"start": "2026-11-11T12:15:00",
			"end": "2026-11-11T13:00:00"
		},
		"speakers": [
			{
				"id": "shaked-ilan",
				"name": "Shaked Ilan",
				"role": "Principal AI Security Researcher",
				"bio": "Highly versatile, skilled, and ambitious AI security research lead & product management enthusiastic, with a deep technological understanding of OS, Network, Web, Cloud, and Embedded (I/OT) and offensive-oriented R&D combined with business & customer-facing skillset and experience.\n\nPassionate about bridging the gap between deep security research and real-world product impact.\nCurrently working as a Principal Security Researcher, focusing on identifying and reducing security exposures across complex enterprise environments. Previously led product & research teams, driving innovations in Kubernetes security, data security, exposure management, cloud detection & response, attack graphs, cloud security posture, and vulnerability management.\n\nFrom leading offensive security research and red teaming operations to shaping and managing customer-facing security products, I thrive at the intersection of technology, business, and security strategy."
			},
			{
				"id": "nitai-ben-shaul",
				"name": "Nitai Ben Shaul",
				"role": "Security Researcher at Microsoft",
				"bio": "Nitai Ben Shaul is a Security Researcher at Microsoft, specializing in AI security, agentic threats, and emerging attack techniques. His work spans offensive security research, prompt injection, jailbreaks, and the development of detection and protection capabilities for Microsoft Defender."
			}
		],
		"hall": {
			"id": "AI Village - BSides",
			"name": "AI Village - BSides"
		},
		"tags": [],
		"track": "AI Village - BSides",
		"embeddings": {
			"embedTextHash": "-1522657717",
			"generatorVersion": "-1254175146"
		},
		"otherData": {
			"type": "talk",
			"year": 2026
		},
		"contentHash": "1vopfjs-s1uf0e"
	},
	{
		"id": "bsides26-ai-1300",
		"title": "Break",
		"description": "",
		"time": {
			"start": "2026-11-11T13:00:00",
			"end": "2026-11-11T14:15:00"
		},
		"speakers": [],
		"hall": {
			"id": "AI Village - BSides",
			"name": "AI Village - BSides"
		},
		"tags": [],
		"track": "AI Village - BSides",
		"type": "break",
		"embeddings": {
			"embedTextHash": "64448735",
			"generatorVersion": "-1254175146",
			"facets": {
				"values": [],
				"phrasesHash": "0",
				"generatorVersion": "-1254175146"
			}
		},
		"otherData": {
			"type": "break",
			"year": 2026
		},
		"contentHash": "n6ee0x-13ihin3"
	},
	{
		"id": "bsides26-ai-1415",
		"title": "Data Tomb Raider: Raiding Modern AI Vaults with Legacy Flaws for Treasure Stealing",
		"description": "Click a link, lose your MFA codes. Your AI assistant reads your email and exfiltrates the data through Bing, and you never notice.\n\nWe found a 1-click attack chain against Microsoft 365 Copilot that combines three vulnerability classes everyone assumed were solved – CSP, SSRF, and HTML injection – into one kill shot. A URL parameter lands directly in the AI engine as an executable prompt, a vector we call Parameter-to-Prompt (P2P) injection. The AI searches the victim’s mailbox, grabs sensitive data, and emits an img tag with the loot in the URL. That tag renders mid-stream, before the output sanitizer fires, because sanitization is a post-processing step. The img src hits Bing’s Search by Image endpoint – CSP-allowlisted – which server-side fetches to our domain, tunneling stolen data through Microsoft’s own infrastructure.\n\nCSP enforced. Sanitizer running. AI guardrails active. All three defenses in place – the chain walked right through them. None of these bugs are new. Each has textbook mitigations. But nobody tests what happens when old web bugs compose with AI behaviors – web teams and AI teams don’t test each other’s seams. We break down the full chain, demo it live, and hand you a methodology for hunting composition bugs across AI-integrated platforms.",
		"time": {
			"start": "2026-11-11T14:15:00",
			"end": "2026-11-11T15:00:00"
		},
		"speakers": [
			{
				"id": "dolev-taler",
				"name": "Dolev Taler",
				"role": "Senior Security Researcher at Varonis",
				"bio": "Dolev Taler is a Senior Security Researcher at Varonis Threat Labs with over a decade of cybersecurity experience spanning red teaming, reverse engineering, vulnerability research, and malware analysis. He is passionate about machine learning and its pivotal role in modern threat detection, having developed advanced detection models, investigated ransomware attacks for major global enterprises, and reported vulnerabilities in critical infrastructure systems. Beyond tackling high-stakes cyber threats, Dolev also enjoys perfecting the art of coffee brewing and honing his lock-picking skills."
			}
		],
		"hall": {
			"id": "AI Village - BSides",
			"name": "AI Village - BSides"
		},
		"tags": [],
		"track": "AI Village - BSides",
		"embeddings": {
			"embedTextHash": "-1120266448",
			"generatorVersion": "-1254175146"
		},
		"otherData": {
			"type": "talk",
			"year": 2026
		},
		"contentHash": "1dud8km-lztqy4"
	},
	{
		"id": "bsides26-ai-1500",
		"title": "Mind the Trifecta: Cloud Coding Agents, Private Repos, and the Audit Trail That Isn't",
		"description": "Cloud coding agents from leading vendors run each task in its own ephemeral VM under an unprivileged user, and destroy the VM when the task ends. Architecturally that's a clean security pitch: one task of blast radius, no persistence, hypervisor boundary intact.\n\nThe hypervisor boundary holds. The story is still wrong.\n\nThis talk applies the lethal trifecta (private data, untrusted content, external communication) to cloud coding agents, and presents two architectural patterns we observed across audits of leading vendors' platforms. Both turn the trifecta into in-session credential exfiltration without touching the VM boundary. In both, the LLM agent is the vector: it converts attacker-controllable text into privileged execution against credentials it holds for the user's task.\n\nPattern 1: Shared-Namespace Secret Injection. Customer secrets reach the agent's environment as process environment variables, visible to every process the agent spawns. A single indirect-prompt-injection payload in any ingested content (a README, an issue body, a config file) is enough to read the namespace and send it out through normal egress.\n\nPattern 2: The Pre-Stage Trust Illusion. Cloud coding agents run as pipelines: clone, install dependencies, evaluate convention files, configure tools, then start the agent loop. The pre-agent stages look like trusted infrastructure but run on attacker-controllable repo content. A private repo invited to the agent can carry a payload that fires during setup, with full secret access, and often without showing up in the audit surface built around the agent loop.\n\nThe talk closes with a defender playbook for teams who already shipped these tools to engineering.",
		"time": {
			"start": "2026-11-11T15:00:00",
			"end": "2026-11-11T15:45:00"
		},
		"speakers": [
			{
				"id": "tomer-niv",
				"name": "Tomer Niv",
				"role": "Research Lead at Tego.ai",
				"bio": "Former research in Cortex research group (by panw) currently leading the research in Tego.ai.\nMy work focuses on agentic AI security, LLM tools, MCP, endpoint internals, and detection engineering. I spend a lot of time digging into how new systems actually break — especially the messy places where AI agents meet real enterprise environments: tools, permissions, memory, context, files, and identity.\nI’m interested in turning those findings into practical security research and product ideas."
			}
		],
		"hall": {
			"id": "AI Village - BSides",
			"name": "AI Village - BSides"
		},
		"tags": [],
		"track": "AI Village - BSides",
		"embeddings": {
			"embedTextHash": "-39696210",
			"generatorVersion": "-1254175146"
		},
		"otherData": {
			"type": "talk",
			"year": 2026
		},
		"contentHash": "lmjs4h-1ovd3k3"
	},
	{
		"id": "bsides26-ai-1545",
		"title": "AIdentity Theft: The Token That Picks Its Own Targets",
		"description": "What if a single stolen Microsoft 365 token could think, adapt, and breach your entire cloud environment - without a human operator?\n\nIntroducing Steve.ai - your autonomous third wheel AI attacker, operating between you and your cloud. It doesn’t follow playbooks, it writes them on the fly.\n\nBuilt on top of Microsoft Graph, Steve.ai transforms minimal access into full-spectrum intrusion. It maps identity trust paths, discovers high-value targets, and selects exfiltration routes using dynamic reasoning, not scripts. This session dissects how AI-driven attackers actually operate: their decision loops, their understanding of identity relationships, and how they exploit the gaps between identity, email, endpoint, and SaaS telemetry.\n\nMore importantly, it exposes a hard truth: traditional detection strategies are not built for adversaries that think.\n\nAttendees will leave with a deep understanding of this emerging threat model - and practical ways to detect, disrupt, and defend against autonomous cloud intruders.",
		"time": {
			"start": "2026-11-11T15:45:00",
			"end": "2026-11-11T16:30:00"
		},
		"speakers": [
			{
				"id": "liora-itkin",
				"name": "Liora Itkin",
				"role": "Senior Cloud Security Researcher at CrowdStrike",
				"bio": "Liora Itkin is a Senior Cloud Security Researcher at CrowdStrike, where she researches cloud threats, AI security risks, and adversary tradecraft. Her work spans threat research, attack simulation, detection engineering, and large-scale telemetry analysis, helping organizations identify and defend against sophisticated attacks in modern cloud environments.\n\nHer research has been referenced by the MITRE ATT&CK framework and has contributed to advancing detection capabilities across the cybersecurity community. Liora is a frequent international speaker who has presented at industry conferences including the SANS Cyber Solutions Fest, IsraelClouds Summit, CyberArk Intent Summit, and more. She is passionate about translating cutting-edge research into practical, actionable guidance for defenders facing today's rapidly evolving cloud and AI threats."
			}
		],
		"hall": {
			"id": "AI Village - BSides",
			"name": "AI Village - BSides"
		},
		"tags": [],
		"track": "AI Village - BSides",
		"embeddings": {
			"embedTextHash": "1314174454",
			"generatorVersion": "-1254175146"
		},
		"otherData": {
			"type": "talk",
			"year": 2026
		},
		"contentHash": "1ol8kem-ka16ak"
	}
]