
AI is changing both the tools available to attackers and the systems businesses need to protect. Understanding that change means distinguishing between capabilities that are technically possible, techniques we can observe in attacks, and risks organisations introduce through their own adoption. Each requires attention, but they do not all call for the same response.
Our analysis of 90 days of phishing across Bridewell’s client base illustrates why that distinction is so important. We found attackers targeting AI assistants, abusing platforms associated with AI development and using AI products as a lure. We did not find evidence of a move towards individually personalised phishing at scale. The findings give organisations specific reasons to act, while challenging assumptions about how AI is being used.
What we can see in phishing
Around 1% of targeted phishing emails in our analysis contained hidden prompt injection. These were instructions intended for an AI assistant summarising or acting on the email, rather than for the person reading it. Microsoft’s prompt-injection protection blocked every identified attempt. The volume was small, but it showed attackers anticipating that an AI system would process their messages.
That is a security consideration for any organisation connecting assistants to its communications. An email can carry instructions aimed at software with access to business information, as well as a lure aimed at an employee. The fact that the observed attempts were blocked is encouraging. It also demonstrates the value of protections designed for this attack route.
We also identified links to phishing pages hosted on platforms including Bolt, Lovable and Replit, alongside developer hosting services such as Vercel and Netlify. Bolt-hosted pages alone appeared in 2% of messages and reached a third of our clients, with half of those emails landing at CNI organisations. AI site builders give attackers a way to produce convincing branded pages quickly, while hosting on legitimate platforms can lend those pages the appearance of trustworthiness.
In 8% of messages, AI products were the pretext. Fake subscription renewals and attachments posing as tools referenced products such as Copilot, ChatGPT and Gemini. This is a familiar social engineering technique taking advantage of a new business priority. An email mentioning AI does not, by itself, tell us whether AI helped create it.
We found no evidence of phishing written by AI and personalised to each recipient. Subject line variation actually fell, with the median dropping from 27 variants per 100 messages in September to 13 by late October. That suggests campaigns were becoming more uniform within this dataset. Subject line variation alone cannot establish whether AI was used elsewhere in producing or running a campaign.
Taken together, these findings show several ways attackers are exploiting AI adoption, without supporting every prediction about highly personalised phishing. Organisations should use that evidence to guide their response. They still need to protect identities, recognise social engineering and investigate suspicious activity, alongside addressing attempts to manipulate AI assistants.
The wider change in capability
Phishing is only one part of the picture. We also see AI used to help build phishing kits, malware and exploits for known vulnerabilities, lowering the technical barrier to those activities. This can make familiar attack methods easier to reproduce and put additional pressure on teams already struggling to identify exposure and remediate it promptly. An attack does not need to be novel to cause serious damage.
Autonomous offensive frameworks
A more consequential shift is now visible in the open. Attackers are no longer only using AI to assist with individual tasks. They are running autonomous, multi-agent frameworks that conduct intrusions with little human direction. ARTEX is the clearest example. On 13 November 2026 the FBI issued a FLASH alert warning that a range of actors had weaponised it, with some intrusions moving from initial targeting to privileged access in under three hours. Microsoft, CrowdStrike, Google and Cisco Talos have each tied it to real intrusions, both state-aligned and financially motivated. India’s financial-sector CERT has named it directly. On 4 November 2026, CSIRT-Fin issued a critical advisory on an AI-assisted campaign it calls Crafty Falcon, in which ARTEX and comparable frameworks are used to abuse payment APIs and move funds. The ConnectWise supply chain compromise shows what this looks like in practice. Microsoft assessed that the China state-aligned actor Storm-0940 used ARTEX with stolen application credentials to attempt sign-ins to more than 2,250 downstream Microsoft 365 tenants.
What ARTEX is matters as much as what it does. It is not malware. It is an open-source penetration testing framework that coordinates specialist agents across reconnaissance, access, privilege escalation and data theft. The autonomy that makes it useful for authorised testing is the same autonomy that makes it dangerous in an attacker’s hands. Frameworks like this lower the skill floor towards zero and operate at machine speed, which shortens the time defenders have to detect and respond.
This changes the economics of testing rather than introducing a new class of vulnerability. If adversaries can run continuous autonomous assessments against exposed applications and APIs, an annual manual penetration test no longer reflects the threat. We reached the same conclusion from the defensive side. We evaluated ARTEX in an isolated lab to find web and API weaknesses before attackers do, and the method transfers directly to our own attack surface. The answer is not more manual testing on a longer list. It is exposure management that moves at the same speed as the threat.
AI also gives defenders useful capabilities. It can help security teams investigate and contain threats faster, and support vulnerability research and testing. Organisations should assess those capabilities against their own weaknesses and operational requirements. Replacing an existing process with an AI tool only represents progress if the resulting capability is demonstrably better.
There is a separate risk in how businesses deploy AI. Assistants are being connected to email, documents and operational processes because those integrations deliver value quickly. Security teams are then left establishing what the tools can reach, which permissions they hold and how dependent the business has become on them. Those questions need answers regardless of how extensively attackers use AI.
An AI system can continue producing plausible, confident answers when its output is wrong or has been manipulated. Where an autonomous agent can act across connected systems, an error can lead to further actions before anyone intervenes. Organisations need to understand the consequences of that access and autonomy, then decide what controls and human oversight each use case requires.
Direct investment towards measurable improvements
Security leaders recognise the issue. Bridewell’s 2026 CNI research found that 39% of respondents ranked managing AI cyber risk among their top cybersecurity concerns, second only to data protection and privacy. Yet that awareness is not consistently translating into additional funding. Some organisations are ringfencing existing resources to investigate the threat, while others have secured one-off investment in capabilities such as continuous penetration testing. Autonomous offensive frameworks turn the shift from periodic to continuous testing from a preference into a requirement.
Incidents can unlock money too. Among respondents who experienced an incident last year, 36% reported a resulting increase in their cybersecurity budget. That is reactive, and the finding does not establish AI as the driver. Businesses need a clearer basis for investment than waiting for an incident or responding to the latest model announcement.
My concern is that organisations, often encouraged by vendors, are replacing capabilities and processes with AI tooling before assessing their overall response to AI risk. A short-term budget increase should give teams the capacity to understand the threat, test useful defensive capabilities and integrate them quickly. Investment should address identified gaps, with evidence that the changes improve protection or response.
Start with capability and control
The starting point is an honest assessment of existing defences. Can the organisation identify and patch vulnerabilities promptly? Do public-facing applications receive sufficient DevSecOps attention? Can teams detect and respond to threats across cloud and operational environments within a timeframe that matches business risk? AI increases the pressure to answer these familiar questions convincingly.
Organisations should measure their current position, define the capability they need and fund the improvements required to reach it. For their own AI adoption, that assessment must also cover the following controls.
Know what you have. Maintain an inventory covering embedded AI features, employee tools and autonomous agents, with clear ownership, data access and process dependencies. Only 29% of respondents in our CNI research managed connected assets centrally and dynamically. That finding concerns connected assets broadly, but it highlights a visibility weakness that AI adoption can compound.
Treat agents as identities. Give each agent only the permissions its defined task requires, and increase autonomy as reliability is demonstrated. Require human approval for consequential actions. Make sure access can be suspended quickly if behaviour becomes unreliable or suspicious.
Make activity observable. Log actions with enough context to establish what happened and investigate the outcome. Test how systems behave when they encounter malicious instructions or manipulated data. Assess whether controls work in the connected business process, as well as in the AI tool itself.
Build and rehearse a fallback. Establish how an AI-dependent process will operate if the system becomes unavailable or its output cannot be trusted. Teams need to be able to recognise that situation and put the alternative into practice. Recovery arrangements should develop alongside adoption.
The evidence should shape both the urgency and the direction of the response. Organisations can strengthen established defences, control the access they give AI systems and use AI to improve security operations now. Their investment decisions should be grounded in what attackers are doing, what their own systems can do and where their defences need to improve.



