New Google Report Exposes 20 Security Vulnerabilities in Various AI Open Source Systems

On August 4, 2025, Google’s Vice President (Security) Heather Adkins announced on X that the company’s AI-powered bug-hunting system Big Sleep had uncovered 20 previously unknown security vulnerabilities in widely used open-source software. Some of these include libraries like FFmpeg and ImageMagick. Big Sleep identified and reproduced each flaw autonomously, and a human analyst later verified the findings before disclosure.

This landmark announcement represents a pivotal moment in cybersecurity, highlighting both the potential and the complexities of applying artificial intelligence to protect open-source ecosystems.

Open Source, Open Risk

Artificial Intelligence doesn’t exist in a vacuum. Most AI tools today, whether powering your smartphone assistant or your enterprise-grade automation platform, are built upon layers of open-source software. From data processing libraries to image manipulation tools, these foundational blocks are often taken for granted as “safe.” But as Google’s Big Sleep project revealed, these blocks are riddled with unseen vulnerabilities.

This presents an ethical crisis as users of these AI systems are unwittingly exposed to risks embedded in the very software that enables AI. Whether you’re using a chatbot, a facial recognition system, or a medical diagnostic AI, your data and outcomes are only as secure as the weakest link in the software supply chain.

Unfortunately, most users have no idea that open-source tools like FFmpeg or ImageMagick might be embedded in the AI products they use. There is no transparency, no disclosure, and certainly no informed consent. Yet when a vulnerability in one of these components is exploited, the user is the one who suffers data breaches, compromised systems, or manipulated outputs.

This lack of informed consent is unethical. This is why AI developers and vendors must be held to higher standards of transparency. Users deserve to know what software is embedded in the products they rely on and what risks that entails.

Potential Risks to Users and Developers

Even with responsible disclosure practices such as Google’s 90-day policy championed by Project Zero, there’s always a risk that vulnerabilities remain exploitable before patches are fully deployed across the ecosystem. Also, the speed enabled by AI may pressure maintainers into hasty fixes. Poorly tested patches can introduce new bugs or compatibility concerns, especially in critical open-source stacks relied upon by millions.

Even the same technology used to detect vulnerabilities could be turned against systems. Adversaries might train AI agents to discover and weaponise flaws at scale, creating a dual-use dilemma.

Disproportionate Harm to Marginalised Users

Security vulnerabilities in AI systems don’t impact all users equally. In sectors like healthcare, criminal justice, or immigration, where AI is increasingly used to make high-stakes decisions, the exploitation of a vulnerability can lead to wrongful diagnoses, arrests, or deportations. These harms often fall disproportionately on marginalised communities.

Ethically, this demands a duty of care from developers and institutions using AI: to prioritise the security of their systems not just as a technical requirement, but as a moral imperative.

What Must Be Done

Mandatory Transparency of Software Stacks
AI companies should be mandated to disclose the open-source components used in their systems, much like nutrition labels on food. Users, regulators, and auditors must know what’s under the hood.
Ethical Auditing of AI Supply Chains
Just as there are supply chain audits in fashion or food industries, we need security and ethics audits of AI’s software dependencies. This is particularly urgent for systems used in critical sectors.
Funding and Support for Open Source Security
Many of the tools Google’s AI flagged are maintained by small, underfunded developer communities. Big Tech companies that profit from AI must ethically contribute to securing these tools, not just scanning them for vulnerabilities but helping to patch and maintain them.
User Notification and Redress
When vulnerabilities are discovered, users of affected AI systems should be notified immediately, and compensation frameworks should be considered for harm caused by these flaws.
Human Oversight and Accountability
While AI can identify threats, only humans can interpret the ethical significance of those threats. AI systems must be governed by robust human oversight structures, especially when they are uncovering flaws that could lead to mass-scale harm.