OpenAI Cybersecurity

Reuben Athaide on why context beats the model in AI security

Reuben Athaide· Head of Innovation, Applications and Assurance at Standard Chartered at Standard Chartered
·~13 min·English·OpenAI
AI SafetyAgentsAI Infrastructure
TL;DR

A Standard Chartered security leader on what changes when frontier AI hunts software exploits: the architecture and trust-boundary context you give the model matters more than which model you pick, and the organizations with solid foundations are the ones that benefit.

01Core Mental Model

Exploitable, Not Just Vulnerable

<strong>Hunt for what is exploitable, not every flaw in the code</strong> - because attackers follow paths across systems, APIs and trust boundaries rather than reading repos line by line.

it isn't really about finding vulnerabilities in code, but it's about understanding what is actually exploitable

— Reuben Athaide, OpenAI Cybersecurity
Key Insight
Counting CVEs measures the code; measuring exploitability measures the attacker. Athaide's reframe suggests remediation effort could follow what is actually reachable rather than the length of the flaw list - a shift in where scarce attention goes, not just how findings are named.
02The Key Lever

Context Beats the Model

<strong>The context you feed the model matters more than which model you pick</strong> - its architecture, dependencies, deployment model and design intent are what surface the findings that are truly exploitable.

it's more important to have that context than which model you pick.

— Reuben Athaide, OpenAI Cybersecurity
Key Insight
If context outweighs model choice, the moat is not access to the best model - most defenders will have that. It is the discipline of maintaining architecture diagrams, dependency maps and deployment intent so any capable model can reason over the real system.
03Earning Trust

Hard Evidence Earns Trust

<strong>A demo does not earn a developer's trust - validated evidence does</strong>, so every AI finding is checked against the code, the architecture, the deployment and the live environment by appsec and developers together.

anything that we find needs to be validated against the code needs to be validated against the architecture deployment and against the environment

— Reuben Athaide, OpenAI Cybersecurity
Key Insight
Athaide seems to treat trust as an engineering output rather than a matter of persuasion: in his telling it comes from validating a finding against the code, the deployment and the environment, which implies a demo that skips that validation may win attention but not adoption.
04Culture Shift

Enabler, Not Gatekeeper

<strong>Security works better as an enabler than a gatekeeper</strong>, building checks into the pipeline and handing developers exploit evidence early so that speed and safety stop being opposites.

it's just trying to change that mindset to be more enablers.

— Reuben Athaide, OpenAI Cybersecurity
Key Insight
The enabler framing reads less as a softer posture than a faster one: a team that blocks at the end tends to create a queue, so moving checks earlier into the pipeline is one way - in Athaide's account - for governance to keep up with AI-speed development rather than fall behind it.
05Prerequisite

AI Amplifies Your Foundations

<strong>Frontier AI amplifies whatever it is given, so the state of your foundations sets the ceiling</strong> - cutting technical debt, limiting language sprawl and tightening identity and access control is what lets a capable model do its best work.

you need to focus on your foundations. Try to make sure that any technical debt is reduced. Reduce any technology complexities.

— Reuben Athaide, OpenAI Cybersecurity
Key Insight
Amplification cuts both ways, so AI widens the gap between well-run and poorly-run organizations rather than closing it. The prerequisite work - cutting tech debt, limiting languages, fixing identity - is unglamorous and is exactly what decides who benefits.
06What Surprised Them

The Gaps You Took for Granted

<strong>The real surprise was stale hygiene they had assumed was handled</strong> - out-of-date architecture and skipped SDLC basics that older tooling had never flagged.

some of the things that you take for granted in the SDLC that you think that development teams are doing like keeping the architecture up to date

— Reuben Athaide, OpenAI Cybersecurity
Key Insight
The surprise here is less a hidden exploit than a hidden assumption - that the architecture docs were current. A context-hungry model effectively audits the hygiene a team believed it already had, which suggests some of what it surfaces may be process gaps as much as code gaps.
07Where It's Going

Secure the Blast Radius, Not the Repo

<strong>Repo-by-repo scanning does not scale for a connected estate</strong>, so defenders need sandboxed bulk runs and a clear picture of the blast radius when one app is breached.

you have a lot of applications that are connected to each other. So you want to understand what is that blast radius

— Reuben Athaide, OpenAI Cybersecurity
Key Insight
Scanning repo by repo secures the parts while leaving the seams between them - the APIs and trust boundaries attackers actually traverse. The 'defense factory' Athaide wants is a bet that security has to model the system as a connected whole, not a pile of independent repos.