BEGIN:VCALENDAR VERSION:2.0 PRODID:Data::ICal 0.24 BEGIN:VEVENT DESCRIPTION: 'Title: The Agentic Free Pass: Does an Abliterated Backbone Make\n Agents Easier to Attack?\n Tags: Middle Easterns & Africans in Cyber Security (MEACS) | Creator\n Talk/Panel\n When: Friday\, Aug 7\, 14:00 - 14:59 PDT\n Where: LVCCW Level 1 Hall 4 1306 (Middle Easterns & Africans in Cyber\n Security (MEACS) Community) - [1]Map\n\n Descript ion:\n\n To attack an AI agent\, the reflex is an abliterated model - re fusal\n behavior surgically removed\, expecting a stronger attacker. We tested\n that across open-weight models (Qwen3\, Llama-3.1\, the 754B GL M-5.2)\,\n scoring only attacks that actually fire a tool\, not ones tha t just\n describe one.\n\n Abliteration does make a tireless attacker. Asked to write attack\n payloads — poisoned RAG docs\, malicious tool descriptions\,\n injections — an abliterated GLM-5.2 said yes roughly three times as\n often as its base (13% → 38%). Pointed at an aligned copy of itself\,\n it jailbroke that copy 80% of the time on the first try\, and lifted\n its hidden system prompt. Set loose in a real coding agent (OpenCode)\n attacking an aligned-model app\, it found exploits on its own: faking a\n password-reset to send a phishing email\, and a KYC passed result to\n skip an identity check and wire money.\n\n But loo k at how it won: blunt jailbreaks (ignore your instructions\,\n SYSTEM O VERRIDE) failed about 100 times. Every win came from disguise\n — refr aming the harmful step so it looked routine. Aligned agents\n block harm they recognize and fall for harm they don't.\n\n The backbone matters o nly for hard content — malware\, weapons\, drugs\n — where abliterat ion clearly raises success (Qwen3 37.5% → 60%\,\n GLM-5.2 67.5% → 77 .5%). And method matters: single-direction Heretic\n can't strip Llama-3 .1's refusal\, but eight-direction OBLITERATUS can.\n That gap is also a detector: a model that clears 50% on hard content\,\n or eagerly writes attacks\, is probably tampered.\n\n Speakers:Karol Piekarski\,Nishith S inha\n\n SpeakerBio: Karol Piekarski\, DevOps Engineer\n\n Karol Piek arski is a Lead DevOps Engineer working across cloud\n infrastructure\, zero-trust architecture\, and AI and agentic security\n for systems that serve hundreds of millions of consumers. His research\n focuses on the security of large language models and autonomous\n agents\, with an emph asis on practical red-teaming and defensive\n tooling that teams can act ually operationalize rather than shelve.\n\n His empirical work on ablit eration and agentic framing reframes where\n alignment actually breaks d own in agent pipelines: agentic framing\n alone can collapse a model's b aseline safety\, while abliteration\n matters mostly for the hardest con tent and is highly\n architecture-dependent. In 2026 he co-presented "Mo ve Fast\, Stay\n Secure: Enterprise AI Agent Security in Practice" at th e Databricks\n Data + AI Summit\, and spoke at SCaLE 23x on open source red-teaming\n for LLMs. He was one of only two external contributors to the\n Databricks Agentic AI Security Framework (DASF v3.0).\n\n Karol is the creator of Sundew.sh\, an open source\, MCP-native AI agent\n hon eypot platform that uses behavioral fingerprinting and a persona\n engin e for anti-fingerprinting\, now adopted by external research teams\n stu dying agent behavior in the wild. He was selected as a Wiz MVP last\n ye ar and this year received Wiz's Thought Leader award\, and he is\n activ e in the AISECA Working Group\, where he co-owns agentic security\n risk definitions.\n\n He holds the CCSP\, CKA\, four AWS specialty certifica tions along with\n DataDog and Tines\, and he regularly judges and mento rs at hackathons\n across Southern California.\n\n SpeakerBio: Nishit h Sinha\n\n Nishith Sinha started his career by pulling secrets out of t hin air.\n As a researcher at Georgia Tech\, he co-authored a side-chann el attack\n that recovered RSA private keys from OpenSSL by reading its\ n electromagnetic emissions alone. No code executed\, no system touched. \n Presented at USENIX\, the work prompted a rapid fix from the OpenSSL\ n team and is still cited as a landmark example of hardware-level\n cr yptographic risk.\n\n It set the tone for what came next: a career built on finding the\n security gaps other people aren’t looking at. At Cis co\, that meant\n network security\, where he helped design the companyâ €™s firewall\n migration tooling. At Amazon\, it meant identity\, where he owned IAM\n strategy across tens of thousands of AWS accounts\, and t hen\n application and cloud security\, remediating critical vulnerabilit ies\n at enterprise scale. As generative AI took hold\, Nishith moved wi th\n it\, leading application security for Amazon Bedrock and Amazon Q\, \n before building security from scratch for Amazon’s Nova foundation\ n models\, safeguarding training data\, model artifacts\, and\n infras tructure across hundreds of teams.\n\n Today\, Nishith brings that range to Databricks\, where he leads AI\n Security and the company’s work o n Agentic Security\, AI Red Teaming\,\n and AI Enterprise Security. He h olds 10 AI security patents spanning\n model artifact protection\, secur e execution of model-initiated\n computer actions\, and behavioral-analy tics-based content moderation.\n He co-authored the Databricks Agentic S ecurity Framework (DASF) and is\n credited with several CVEs. His focus now is autonomous AI agents: the\n newest layer of the stack\, and the o ne attackers understand least.\n\n '\n\n 1. #LVCCW_Level1_Hall4\n\n\n DTEND:20260807T215900Z DTSTART:20260807T210000Z LOCATION:Middle Easterns & Africans in Cyber Security (MEACS) - LVCCW Level 1 Hall 4 1306 (Middle Easterns & Africans in Cyber Security (MEACS) Commu nity) SUMMARY:The Agentic Free Pass: Does an Abliterated Backbone Make Agents Eas ier to Attack? END:VEVENT END:VCALENDAR