Identity verification has been the core of cybersecurity with one assumed truth: verify identity at the point of entry, then trust the session. The new wave of AI has weakened this faith, operating at machine speed to mimic behaviour and bypass controls, meaning identity must be continuously monitored and re-verified. With multi-step AI-assisted attacks rising 180% year-on-year in 2025 according to the World Economic Forum, this is something we see leaders in the cyber space reacting to quickly.
1. Continuous verification in Zero Trust architectures
Rather than assuming trust persists after authentication, Zero Trust architecture rejects the traditional “castle-and-moat” model of securing the perimeter and trusting everything inside it. Instead, it divides an organisation’s IT infrastructure into smaller, independently protected segments, each with its own verification requirements. This means that even if an attacker compromises one set of credentials, they face new authentication barriers at every boundary. Strong identity verification at entry remains important, but the main change is how the blast radius of compromised credentials is significantly reduced.
Compromised credentials are one of the most common causes of a breach. IBM’s 2025 Cost of a Data Breach Report shows breaches linked to stolen credentials cost an average $4.81m and take over 290 days to detect and contain. They also stated that 93% of organisations experienced 2 or more identity-related breaches in the past year, highlighting the severity of the threat.
2. Widening attack surface
As organisations use more applications, operate increasingly in the cloud, and work remotely across multiple devices, the ways that their employees can be targeted are multiplying. AI is accelerating this by improving the quality of targeting through deepfakes and hyper-personalised phishing attacks, which make social engineering attacks more convincing than ever. Humans are a massive risk, which is why limiting devices and upskilling people through training and phishing simulations remains essential.
But alongside this, a second and distinct attack surface is emerging: the non-human one. AI agents can now autonomously identify what to access, determine the optimal moment to do so, and chain actions across systems without any human in the loop. This is very different from the risk profile of automation scripts in the past. As these agents act across customers’ environments, cyber suppliers will increasingly be judged not just on whether an identity was verified at the start, but on whether loss of control was managed throughout: whether access was appropriately decayed over time, whether privileges were narrowed as risk increased, and whether the system could intervene mid-execution. Both perimeters are expanding simultaneously.
3. The next wave of identity shocks
Now, whether you believe that Mythos is as dangerous as Anthropic has outlined, or if it’s just a clever marketing ploy, it is clear that as AI models get more sophisticated, so will the imminent danger to any cyber vulnerabilities. Bain characterises Mythos as a ‘signal rather than the threat itself‘, which will expose under-investment in foundational cyber controls. A clear risk is how organisations govern the identity and access of their own AI agents. As businesses deploy copilots and autonomous agents across their operations, those agents need to be subject to the same identity verification and access controls as the people they work alongside. An agent that can access payroll data, financial records, or sensitive customer information when the human equivalent would be denied that access entirely — represents a significant governance gap. If not architected correctly, your own tooling can become a source of vulnerability.
This risk is compounded by an accelerating asymmetry in how attacks are developed and deployed. AI is enabling attackers to release malware at a scale and speed that was previously impossible, with little need for quality control. Cyber attackers have the privilege of pushing beta versions and seeing what works to iterate rapidly, while defenders need to rigorously test and ensure the solution can work across complex interconnected systems. AI widens this gap further, which is why any improvement in model capabilities further expands the exposure of vulnerabilities. There is an arms race underway, and those not investing in the foundational controls now risk being left behind.
Move beyond Mythos or any other next-gen model, and on the horizon is quantum computing. While it’s likely a decade away, companies are already adopting encryption algorithms that can withstand the onslaught of a quantum attack. In fact, Europe has already set a 2035 deadline for removing quantum-vulnerable cryptography, with some governments targeting 2030 for sensitive systems. While 2035 may feel like a long time away, there is growing concern around cyber attackers adopting a “harvest now, decrypt later” approach to collect encrypted data now, betting on future technology to decrypt and use it in attacks, meaning companies need to identify higher-risk encrypted data now and move to protect it ahead of time.
