Hi, I'm Tara! I'm a CS student at Stanford interested in security, especially where the security we expect on paper breaks down in practice. Recently, I've been working on malware-model reliability, vulnerability intelligence, enterprise security, and security for autonomous AI systems.

Currently → building ContainmentBench to study security controls for autonomous coding agents.

Research

What happens when the reassuring result isn't the whole story?

AI Security

ContainmentBench

In progress

I started building an evaluation environment for testing whether security controls could contain autonomous coding agents across multi-step Linux tasks. The project focused on a failure mode that single-step checks can miss: individually allowed actions can become unsafe when they add up over a longer trajectory.

Question Are individually safe actions still safe when an agent combines them across a longer trajectory?

Empirical Security · Model Reliability

Hidden Reliability Failures in Malware Detection

I evaluated EMBER2024's released malware detector on 1.08 million Windows PE files. Its overall numbers looked excellent (98.0% accuracy and 0.0031 ECE), but family-level analysis found malware families with false-negative rates above 80%. The results showed how strong aggregate metrics can hide concentrated security failures.

Expected 98.0% accuracy · 0.0031 overall ECE
Found 85.2% false-negative rate for one malware family

Vulnerability Intelligence · ML

Predicting Vulnerability Attributes from CVE Descriptions

Inspired by NIST's struggle to keep up with record CVE growth, I tested whether models could fill in the vulnerability details defenders need to prioritize newly disclosed CVEs. Using only CVE descriptions, I predicted severity, CWE, and CVSS attributes. On 36K+ held-out 2024 CVEs, traditional text models (TF-IDF) beat zero-shot 70B+ language models on severity and CWE prediction. Rare classes and details that were implied rather than stated directly remained hard across the benchmark. The results showed that bigger models were not automatically better at this task.

Expected 70B+ language models should have the advantage
Found TF-IDF beat zero-shot 70B+ models on severity and CWE

Security Engineering

Compliance Reality Checker

I built Linux and Microsoft 365 checks that compared declared security policy with account-level exceptions and runtime state. The checks caught cases where a control looked enabled at the configuration level but was not actually enforced for every user or system.

Looks secure Control is configured
Reality Configured ≠ enforced

Publications

Empirical Security

Investigating Usefulness of Security Threat Reports

Author · Manuscript under review

We compared 31 incident reports from DFIR firms and cyberinsurers and found surprisingly little agreement on basic questions like the most common initial-access vectors or most-exploited CVEs. Inconsistent definitions, missing sample sizes, and unclear methods often made the disagreements difficult to interpret. The bigger problem was not just conflicting numbers, but how hard it was to tell which numbers were trustworthy.

Expected Major incident reports should converge on basic threat trends
Found No CVE appeared in all five comparable most-exploited lists

Empirical Security

A First Look at Governments’ Enterprise Security Guidance

Research contributor · USENIX Security '25 · August 2025

The study compared government enterprise-security guidance across 41 countries. In a deep comparison of 10 frameworks, only 2 of 166 observed controls were recommended by all 10, and even close security allies disagreed substantially. The results raised a broader question: how much of security “best practice” is actually backed by empirical evidence?

Expected Security “best practices” should broadly agree
Found Only 2 of 166 observed controls were universal

Paper

Privacy · Technology Ethics

Privacy and Paternalism: The Ethics of Student Data Collection

Author · MIT SERC · August 2022

We examined how school monitoring technology changed student privacy and autonomy. The case study asked when safety-oriented data collection becomes too intrusive, especially when students have little control over how their data is collected or used.

Paper

Contact

I'm always interested in meeting people working on AI security or related problems. Feel free to email me at tdixit@stanford.edu.