New opportunity

AI Researcher: Adversarial Systems

Auric AI · Bengaluru, India

About this role

Auric AI seeks an AI Researcher: Adversarial Systems to independently challenge and evaluate an agentic reasoning system used for intelligence and defence work in India. The role focuses on finding convincing but incorrect conclusions, testing the assumptions behind them, and establishing whether the system and its safeguards work. The posting lists 0–10 years of experience, says no seniority or defence background is needed, and specifies on-site work in Bengaluru or Mumbai.

Responsibilities

Develop adversarial approaches that find alternative explanations, test consequential assumptions, and identify evidence missing from apparently sound conclusions.
Attack the system through poisoning, injection in material it reads, saturation and probing; build defences and demonstrate that they hold.
Develop evaluation methods using held-out ground truth, scoring by question class, variance across runs and drift detection, particularly where reviewing every output is impractical. Distinguish critiques that reveal real flaws from merely plausible objections.
Design trials to detect whether fluent machine assessments anchor users and degrade human judgment despite apparent gains in throughput.

Relevant background and approach:

Useful backgrounds include adversarial machine learning, security research, multi-agent systems, reinforcement learning, game theory, forecasting and calibration, causal inference, and experimental design. Grounding in statistics or philosophy of science alongside engineering is particularly valuable.
The role calls for rigorous scepticism, a preference for experiments that can disprove claims, and the ability to argue for a flaw against resistance. The system uses messy multi-source data and self-hosted models, without fine-tuning; humans retain final judgment. The posting cautions against assuming self-play results transfer to this partially observable domain with adaptive adversaries and delayed or unobservable outcomes.

To apply, send a resume and an argument of no more than two pages identifying and challenging a key assumption in the proposed system to careers@auricai.in.

For full details about this role, click “Apply on company site” to view the employer’s original posting.

Skills for this role

Adversarial machine learningSecurity researchMulti-agent systemsReinforcement learningGame theoryForecastingCalibrationCausal inferenceExperimental designStatisticsPrompt injectionData poisoningDrift detection

Checking your profile…

Preparation suggestions, not the employer’s actual interview questions. Always check the original posting for current requirements.

Understand the role

What matters most

Find consequential reasoning failures
The role independently challenges conclusions that may be fluent, well-cited and wrong. It calls for alternative explanations, tests of pivotal assumptions and attention to evidence that should exist but is missing.
Attack the system and verify defences
The posting names poisoning, injection in material the system reads, saturation and probing. The work includes building defences and demonstrating that they hold, not just producing attacks.
Build evidence of reliability
Because every output cannot be reviewed, the role calls for held-out ground truth, scoring by question class, variance across runs and drift detection. It also requires distinguishing critiques that expose real flaws from plausible-sounding objections.
Test effects on human judgment
The role includes designing trials to detect whether fluent machine assessments anchor users and degrade their reasoning, even when apparent throughput improves. Humans retain final judgment in the system described.

Likely challenges to think through

Separate genuine flaws from persuasive critiques
Preparation hypothesis: An objection can sound rigorous without changing whether a conclusion is wrong. Prepare a way to test a pivotal assumption, specify what evidence would discriminate between explanations and explain what result would weaken the critique.
Evaluate without a dependable reward signal
Preparation hypothesis: Partially observable conditions, adaptive adversaries and outcomes that may remain unknown for years make simulator or self-play success hard to trust. Prepare to discuss what held-out cases and intermediate measures can establish—and where they cannot establish real-world correctness.
Preserve independent criticism while testing useful defences
Preparation hypothesis: A critic needs enough system access to probe attacks and verify defences, but shared assumptions could reproduce the evaluated team's blind spots. Prepare an approach that makes attack results reproducible while testing conclusions against alternative explanations.

Practice your answers

Technical practice

How would you test an assessment that is fluent, well-cited and internally coherent but wrong?

What to coverPreparation suggestion: Trace each consequential claim to its cited evidence and check whether the evidence supports the inference, not just the wording. Explain how you would test alternative interpretations and record uncertainty where the sources cannot settle the question.

How would you find the assumption doing the most work in a conclusion?

What to coverPreparation suggestion: Identify which assumption would most change the conclusion if false, then propose a test that could disprove it. Look for evidence that should exist under the conclusion but is missing, while accounting for incomplete multi-source data.

How would you test poisoning and instructions embedded in material the system reads?

What to coverPreparation suggestion: Distinguish corrupted source content from an instruction attempting to redirect the system, and describe controlled tests for each. Assess whether the resulting conclusion or cited evidence changes, not merely whether the suspicious text is repeated.

What would saturation and probing reveal about the system?

What to coverPreparation suggestion: Describe how repeated, varied inputs could expose what the system reacts to or cause it to miss consequential evidence. Explain how you would measure both answer quality and information revealed to an adaptive adversary.

How would you demonstrate that a defence holds after discovering an attack?

What to coverPreparation suggestion: Retest the original failure and variations that the defence has not seen, using explicit failure criteria. Discuss the trade-off between blocking attacks and preserving the system's ability to use legitimate source material.

How would you evaluate correctness when reviewers cannot inspect every output?

What to coverPreparation suggestion: Explain how held-out ground truth and scoring by question class could reveal failures hidden by an aggregate score. Address how you would select cases and interpret results when reliable ground truth is limited.

How would you detect unstable performance or drift?

What to coverPreparation suggestion: Compare repeated runs on the same cases and track results by question class over time. Distinguish run-to-run variance from a sustained change, and explain what evidence would justify investigating either.

How would you distinguish a real flaw from a plausible-sounding critique?

What to coverPreparation suggestion: Require a critique to identify a consequential claim, specify what would falsify it and connect the objection to evidence. Explain how you would handle objections that sound persuasive but cannot yet be verified.

Why might self-play results fail to transfer to this domain?

What to coverPreparation suggestion: Address partial observability, adaptive adversaries, an uncertain simulator and outcomes that may remain unobservable for years. Discuss what limited evidence could still be useful without treating a simulated reward or equilibrium as proof of real-world performance.

How would you test whether machine assessments improve throughput but worsen human judgment?

What to coverPreparation suggestion: Propose a controlled comparison of decisions made with and without the assessment, measuring reasoning quality as well as speed. Account for anchoring on fluent but incorrect outputs and for humans retaining final judgment.

Behavioural practice

Tell me about a time you challenged a conclusion others found convincing.

What to coverPreparation suggestion: Use a real situation, your action and the result. Focus on the assumption or missing evidence you identified, how you made the case against resistance and what the evidence ultimately showed.

Describe a time you designed a test that could disprove your own preferred idea.

What to coverPreparation suggestion: Explain the situation, the falsification criterion you chose, the action you took and the result. Be clear about how you responded if the outcome was inconclusive or contradicted your expectation.

Give an example of evaluating work while maintaining an independent perspective.

What to coverPreparation suggestion: Describe the situation, how you avoided inheriting the team's assumptions and what you did to test your concerns fairly. State the result without presenting disagreement itself as evidence of a flaw.

Tell me about a time you communicated a consequential uncertainty to decision-makers.

What to coverPreparation suggestion: Set out the situation, what was and was not known, the action you took to communicate it and the result. Explain how you avoided overstating confidence while leaving the final judgment with the people responsible for it.

Sharpen your knowledge

Topics to prepare

Alternative explanations and missing evidence
Posting focus: challenge fluent, well-cited conclusions by testing consequential assumptions and noticing evidence that should exist but does not. Revision exercise: take a fictional multi-source assessment, write two competing explanations, and specify one observation that would weaken each.
Limits of self-play and simulation
Posting focus: the company rejects straightforward transfer from self-play because observations are incomplete, adversaries adapt, and outcomes may be delayed or unobservable. Revision exercise: explain which conclusions from a simulated trial would remain uncertain if its rules or reward signal were wrong.
Evaluation under incomplete review
Posting focus: held-out ground truth, scoring by question class, run-to-run variance, and drift detection when every output cannot be inspected. Revision exercise: sketch how a fixed set of known-answer cases could reveal a failure hidden by an aggregate score.
Real flaws versus plausible objections
Posting focus: distinguishing a critique that exposes an actual flaw from one that merely sounds convincing. Revision exercise: write three objections to a fictional assessment, then state what evidence or test would substantiate each.
Human judgment and anchoring
Posting focus: fluent machine assessments may increase apparent throughput while degrading users’ reasoning. Revision exercise: outline a comparison between judgments made with and without a machine assessment, including a measure of decision quality rather than speed alone.

Skills to revise

Design a controlled adversarial test
Posting focus: poisoning, injection in material the system reads, saturation, and probing. Suggested practice: use fictional documents to define one attack, its intended failure, a control case, and the observations needed to tell whether it worked.
Test whether a defence holds
Posting focus: build defences and demonstrate their effectiveness. Suggested practice: for the fictional attack, specify a mitigation and rerun both attack and control cases, recording remaining failures rather than declaring success from one blocked example.
Construct a measurable evaluation plan
Posting focus: assess correctness without reviewing every output. Suggested practice: define held-out cases, question classes, repeated runs, and a drift check; explain how you would report uncertainty and class-specific failures.
Write a falsifiable challenge
Posting focus: the application asks for an argument of no more than two pages attacking a key assumption of the described system. Suggested practice: draft a concise claim, the assumption it challenges, a disconfirming test, and what result would change your view.

Your preparation checklist

Draft the required argument
The posting asks for a maximum two-page argument challenging a key assumption. Choose one assumption about reasoning over messy multi-source data, state how it could fail, and propose a test that could disprove it.
Prepare a project walkthrough
Select relevant work, if available, in adversarial ML, security research, evaluation, or experimental design. Prepare a concise account of the hypothesis, method, evidence, limitations, and what changed when results challenged expectations.
Sketch an attack-and-defence exercise
As interview preparation, use a hypothetical case to show how poisoning or an injection in source material could produce a convincing but wrong conclusion. Specify the proposed defence and the evidence needed to show it holds.
Build an evaluation plan
Outline a sample evaluation artifact using held-out ground truth, scores by question class, variance across runs, and drift checks. Include a way to distinguish a critique exposing a real flaw from a merely plausible objection.
Design a human-judgment trial
Propose a controlled comparison of decisions with and without a fluent machine assessment. Specify how to detect anchoring or degraded reasoning, not just increased throughput.
Prepare questions about independence and evidence
Ask how the adversarial function remains independent, how evaluations work when outcomes are delayed or unobservable, and what evidence would persuade the team that a safeguard failed.
Check the application materials
Pair the resume with the two-page-or-shorter argument requested in the posting. Check that both address this role before sending them to the stated application email.

Availability last checked: 4 Nov 2026. Status can change after a check; confirm on the employer’s website. Report an issue · Our checking policy

KEEP EXPLORING

Similar AI jobs

Related skills and specializations in India. Matched to this role, not your profile.

Explore more jobs

Finding similar opportunities…