Source: The Conversation – France in French (2)– By Christine Abdalla Mikhaeil, Assistant professor in information systems, IÉSEG School of Management
The American company Anthropic, specialized in generative artificial intelligence, has takenearly Aprilthe decision to freeze the public release of one of its recent models, named“Mythos”.
According to a company statement, this decision stems from a computing power and reasoning capacity deemed too “offensive.” Anthropic has chosen to share its model only with a coalition of tech giants (Apple, Amazon Web Services, Cisco, Google, Microsoft, etc.) within the framework of the Glasswing project. The declared goal is to use Claude Mythos Preview to detect so-called vulnerabilities “.zero-day” (that is to say unknown and having no known fix) andproactively secure critical software… before malicious actors exploit these vulnerabilities.
Large language models have been able to code for several years now, but the specialized press is now documenting a more worrying leap. Artificial intelligence (AI) systems canidentify vulnerabilitiesreal in critical software. Authorities, such as the National Cybersecurity Agency of Information Systems (Anssi),emphasizethe ability of AI systems to automate attacks.
The stakes of the widespread dissemination of such models, including Mythos, go far beyond the technical scope. A large-scale cyberattack, automated by AI, could paralyze financial or logistical systems within seconds, with a remediation cost running into billions of euros. The stakes are also societal and health-related, since our hospitals, energy networks, and other critical systems rely on often outdated software layers, vulnerable to “attacks fromzero-day“from now on generated on the chain.
In this context, can ultra-powerful AIs, like Mythos, contribute to a form of “algorithmic deterrence”? This is based on a simple principle: detecting and neutralizing one’s own critical vulnerabilities faster than any human or automated attacker – including during an attack – so quickly that the attack becomes useless or too costly.
The withholding of this model by one or more private American companies also reopens the question ofdigital sovereigntyAt the global level.
AI systems facilitate cyberattacks
Historically, cybersecurity is based on a fundamental asymmetry: the attacker only needs to find a single security flaw, while the defender must fix them all in a kind of race against time.
The integration of AI systems enhances attackers’ capabilities, primarily because a model like Mythos can scan millions of lines of code in a few minutes, whereas a human would spend weeks analyzing the source code of software to detect a memory error. This is what is called ‘automation of recognition.’
Moreover, AI enables thephishingof high precision, that is to say fraudulent messages (thephishingclassic) but more credible, without spelling mistakes, in any language and ultra-personalized to deceive the user. TheAnssi also warns about the use of generative AIto break down the traditional linguistic and psychological barriers that make the “human firewalls” – that is, the vigilance and critical spirit of readers – increasingly obsolete.
Finally, somemalwarecan now rewrite their own code in real time to evade detection “by signature”. This classic antivirus method consists of identifying a virus by its “fingerprint” (an already known and recorded code). By constantly changing form (polymorphic exploitation), themalwarebecomes invisible to these traditional tools.
AI for cyber defense
Conversely, AI also improves cybersecurity capabilities, thanks tocausal analyses, which allow modeling the relationships between events, as well as accelerating the identification of anomalies by aautomated monitoring and the prioritization of their corrections. Thus, an AI analysis system enabled in January todiscover 12 security vulnerabilities in OpenSSL, essential software for protecting global internet communications.
Mythos also seems already to participate in this automation, andFirefox already claims to have identified and fixed 271 vulnerabilities thanks to this software, which suggests that Mythos indeed excels at vulnerability detection when it has access to the source code.
On the other hand, nothing currently proves that Mythos can, without access to the source code and without human intervention, autonomously compromise any closed software.
Moreover, someanalysessuggest that comparable capabilities could already be replicated from public models, calling into question the effectiveness of this retention. Thus, Mythos today resembles more a powerful security analyst, capable of identifying vulnerabilities and proposing exploitation avenues, than an autonomous universal cyberattack entity.
What really worries is not just that Mythos knows how to code better or test code better: it is that it seems to lower the cost, time, and level of expertise required to discover and chain vulnerabilities, thus potentially accelerating both defense and attack.
Towards a new balance of “algorithmic deterrence”?
In this context, the notion of “algorithmic deterrence”algorithmic deterrencein English) emerges. It can be understood by analogy with nuclear deterrence: it would no longer be just about protecting oneself, but about possessing a response and detection capability so rapid that the attack becomes useless or too costly.
Unlike the nuclear domain, algorithmic deterrence relies on strengthening defensive mechanisms (rather than response): accelerated intrusion detection, cause analysis, and attack simulation to plug vulnerabilities before they are exploited.
Before the era of AI, algorithmic deterrence was more limited: security teams carried outpenetration tests, a security assessment method that relies on simulating cyberattacks to identify open ports and known vulnerabilities, in order to trigger their correction.
Today, we saw it,AI can strengthen defensive mechanismsand therefore deterrence. But, in the best case, a defensive AI allows for reducing the cost of protection and increasing the probability that an attacker is detected or neutralized before reaching their objective.
Algorithmic deterrence therefore remains fragile. Even in the era of AI, it depends heavily on the practices of operational actors (cybersecurity agencies, armies, companies), the quality and modernity of the inherited systems they must protect and integrate, of thenational and military strategies implemented by Statesas well as devices forgovernancewhich define the rules, responsibilities, and control mechanisms of AI.
The dilemma of retaining deterrence tools: security versus transparency
Not making certain models available to the general public may seem responsible, as it avoids publishing offensive capabilities that could be immediately misused. But this withholding concentrates technological power in a few hands and reduces scientific transparency.
The debate here converges with that of the frameworkthe European AI Act, which already imposes obligations oftransparency, traceability, and documentationfor general-purpose AI models, while seeking to reconcile innovation, security, and protection of industrial secrets.
The opacity of AI models limits their auditability, hinderingdevelopment of appropriate countermeasuresandconcentrates technical power among a few mainly American actors, to the detriment of open research and democratic governance. This criticism is part of a broader academic literature showing that the opacity of artificial intelligence systems compromises theirreproducibility, their auditability and,ultimately, theirscientific value.
Christine Abdalla Mikhaeil is a member of the Association for Information Systems (AIS). She has received funding from the Bachelier Institute and the LEM CNRS UMR 9221.
–ref. Algorithmic deterrence, retention: is AI ushering cybersecurity into a new era? –https://theconversation.com/algorithmic-dissuasion-retention-does-ai-move-cybersecurity-into-a-new-era-282192