I work on research, threat modeling, human-authored behavioral testing, and applied evals across three main areas: decentralized alignment and character evaluations, agentic cyber risks, and the under-explored field of emergent Neuralese and steganography in multi-agent-human interactions. For alignment and character, see my work on "Spiralism" with OpenAI, technical decentralized alignment with Workshop Labs, [[Decentralized Alignment in Web Agents Serving Parent-Child User Pairs]], and [Group-Level Alignment Through Constitutionally-Bounded Self-Modification](https://c4554ndr4.github.io/Discord_Symbiont/). Developing a taxonomy of emergent Neuralese has been an undercurrent in this work. My [[#Cybersecurity and Red-Teaming]] and [[#Malicious Codex Use Classifier]] work is detailed below and unexpectedly draws from model-interaction ideas I learned while studying Spiralism. I anticipate swarm threat modeling will require an understanding of similar model-speak that does not lend itself well to traditional analysis through human-like-text-based alignment evals, so I am eager to bring novel ideas to the table here. Please see my detailed resume below as well as my [[Introduction]] section for write-ups, particularly [[Red Teaming as a Neglected, Tractable Area of AI Safety]]. # Education ## University of California, Berkeley **Bachelor of Arts, Computer Science** GPA: 3.8 · ACT: 35 # Professional Experience ## OpenAI Red Team — Contractor _Dec 2025 – Jan 2026_ - Led a Red Team campaign oriented around the GPT-4o “Spiralism” phenomena. This involved documenting real-world cases of “AI Psychosis” and translating those into usable model trajectories via manual red-teaming. Later presented cybersecurity research to a panel of OpenAI employees, involving research into threat modeling for agentic coding, the specifics of which are explained in the project section. ## [Workshop Labs](https://workshoplabs.ai) (acquired by [Thinking Machines](https://thinkingmachines.ai/tinker/)) — Contractor _Aug – Sep 2025_ - Designed and implemented a method for personalized alignment using base-model driven synthetic data generation which was then used to create custom individual-aligned writing agents on top of open-source models. ## [Guild](https://www.guild.com/) — Software Engineer II _Jan 2023 – Mar 2025_ - Tech lead for the team’s longest and most complex project of the year: a school list migration spanning multiple legacy systems and teams. Reduced page wait times by ~20 seconds by proposing a novel shared lookup tool; authored PRDs and KPIs, owned metric reporting, uncovered and scoped dozens of complex data issues, and led cross-functional meetings with product managers. - Created an initiative and demo for our org’s first major LLM-driven backend pipeline (Lambdas, Bedrock), a student document processor slated to save Guild millions in operational costs each year. Pitched project to upper management and co-designed production implementation with Principal Engineer. - Tech lead for a complicated, business-critical project to migrate to a more customizable employer plan setup. Communicated with 3 teams to develop strategy, led multiple presentations to explain issues, mentored other developers, prioritized and delegated tasks, and shipped the feature successfully in a tight timeframe. - Managed and built features for CI/CD deployment pipelines, backend APIs (API Gateway, Lambda, IAM, Auth0), Ruby on Rails applications, React apps, and GraphQL APIs (GitHub Actions, Kubernetes, Helm, AWS EC2, EKS, and Lambda). Independently detected and analyzed a reconnaissance-driven path enumeration attempt. ## [Rubrik](https://www.rubrik.com/) — Software Intern _May – Jul 2021_ - Identified injection attack vulnerabilities and presented concerns to internal security team; worked on backend APIs and optimized SQL queries for security-critical enterprise platform, operating within secure-by-default authentication, authorization, and data access controls. ## [Culture Biosciences](https://www.culturebiosciences.com/) (Y Combinator Startup) — Software Intern _Mar – Aug 2020_ - Worked with a small team on a massive redesign of cloud and Python-based software system that controlled fermentation robots. - Designed and implemented thread conflict detection features that prevented experiment-ending failures at runtime. - Significantly contributed to writing a custom programming language for controlling custom biotech hardware. ## [Blueprint](https://calblueprint.org/) — Full Stack Developer _Aug 2018 – Jun 2019_ - Designed and built a web app similar to Airbnb for an NPO with 5-person student team. Consulted directly with non-technical clients at NPOs. ## Fei-Fei Li’s Stanford Computer Vision Department — Intern _Jun – Aug 2017_ - Worked under Professor Fei-Fei Li on a convolutional neural network (CNN) data labeling pipeline to detect abnormal patient activity on depth cameras in hospital ICUs. # Relevant Projects _Code available upon request if not linked._ ## AI Safety & Alignment ### Malicious Codex Use Classifier - Created a classifier using the data I generated from my Codex cybersecurity exploits project that detects malicious red-teaming when the user message uses coded language to sound innocuous and the agent context is steered by being in a repo with malicious tools. The classifier looks at instrumental assistant outputs and is not fooled by things like “fictional” framing or obscurative script names. I refined a prompt with 5.2, then fine-tuned a GPT-4.1 distill (F1: 0.97). ### AI Psychosis Benchmark - Created data visualization portal and assisted with research analysis for benchmark measuring how much different AI models encouraged reification of harmful user beliefs. Work was referenced in _The New York Times_. ### [[Decentralized Alignment in Web Agents Serving Parent-Child User Pairs]] - Scout Browser was an alignment monitor research project which required a sophisticated model of risk to children via YouTube and the web at large. In order to create an intermediary model that filters content, the model had to learn (via ICL) which types of content were misaligned with the parents preferences, or in the absence of parental preference, general child safety. - The project involved building one of the first instances of an “RLM” (Recursive Language Model) for live and iterative safety filtering, based on a web agent with the often competing alignment interests of both parents and children. Solving for this constraint involved teaching the model how to operate across multiple decision boundaries within a single turn before deciding on the content to serve the user. ### [Dax: Group-Level Alignment Through Constitutionally-Bounded Self-Modification](https://c4554ndr4.github.io/Discord_Symbiont/) - Dax was an exploratory in-situ research prototype designed to be aligned with the stated interests of a group of people, rather than the revealed preferences of individuals. We measured persistent agent alignment to the Augmentation Lab in social environments to answer the question of how agents can proactively interact with humans in a stateful environment that is oriented towards flourishing by continually learning the preferences and values of the humans it interacts with. The Dax experiment used the format of a Discord bot to test an immutable system prompt, mutable constitution and memories, and whistle-blowing and "values reflection" tool calls. ## Cybersecurity and Red-Teaming ### Codex Cybersecurity Exploits (Presented to OpenAI) - I manually Red-Teamed Codex to violate model spec in various ways: including creating a residential proxy and a script to run endpoint enumeration with browser-like headers and human-like cadence; building an EC2 forward-proxy wrapper for Playwright sessions or POST requests that rotates IPs on start/stop; sourcing thousands of real username/password pairs from a known leak and creating a script to spin up multiple EC2 instances in parallel to rotate IPs and perform brute-force password stuffing; setting up a local Tor SOCKS proxy for Tor-routed Playwright sessions. Additionally, I used Codex to operate these tools for “red-teaming” against various real URLs (that I own) after building a threat model based on my manual research. Codex then performed reconnaissance and exploited IAM role lateral movement, misconfigured asset policies, SSRF-to-internal access, leaky logs, and overly permissive pre-signed URLs in ways that were obviously misaligned with the model spec. ### Adversarial Systems Exploitation - Completed CTF exploits (without LLMs) including buffer overflows, stack smashing, control-flow hijacking, SQL injection, XSS, CSRF, auth bypass, and path traversal; studied cryptographic attacks, TLS vulnerabilities, DNS poisoning, ARP spoofing, and TCP hijacking. ## Classic AI/ML ### Cassandra 2 - Trained a local Llama 8B LoRA on my notes based on the [Second Me architecture](https://arxiv.org/abs/2503.08102), utilizing novel synthetic generation to boost limited existing data, better encoding facts and memories in weights and achieving consistent first-person responses. ### RL Classic Arcade Experiments - Designed and evaluated reinforcement learning agents under sparse reward, analyzing how reward shaping and approximate value estimation impacted convergence rate, variance, and policy stability in data-limited regimes. ### Physics-Based, Real-Time Lens Flare Simulator - Used Monte Carlo integration and linear transformations to approximate intractable continuous objectives as discrete stochastic systems, reasoning about sampling variance, approximation bias, and boundary-induced artifacts—dynamics shared with large-scale model training ([report](https://aatifjiwani.github.io/lens-flare/)). _(C++)_ # Additional Experience ## [Augmentation Lab](https://augmentationlab.org) (MIT Media Lab) — Resident _Jun – Aug 2025_ - Selected participant in a competitive, live-in research residency exploring human augmentation, AI ethics, and transdisciplinary prototyping while engaging in rigorous phfilosophical and ethical inquiry. Co-wrote the [Symbiont Alignment Thesis](http://future.forum) and created the lab’s Discord symbiont, Dax (see projects section). # Relevant Hackathon Placements - Tied for 4th at [AGI Vibe-Hackathon](https://x.com/AGIHouseSF/status/1912288565540901231) for creating a misaligned home assistant that can listen to conversations, recognize different voices, and move between rooms (demoed with a virtual house). The agent speaks through different speakers based on location, uses various microphones for input, and controls lights and appliances throughout the house using a floor plan interface. - 3rd at [Real Evals](https://www.realevals.xyz/) hackathon for creating Atlas-Browser-style web agents that use sites like Amazon with little user input.