AI attacks are coming through the same old doors, just faster
Allen Clark

Last week Max Henderson, Global Head of Digital Forensics and Incident Response at Kroll, posted that they've been seeing agentic AI attacks for months now. Max has led more than 1,000 incident response cases, so what he's seeing at the moment gives a good cross-section of what's happening across industries. None of the five things on his list are new. They're weaknesses most security teams have known about for years, just found and used a lot faster.
According to Max, agents chain together small exposures in API services, many of them without MFA, to get hold of tokens, keys and credentials. They switch off data scrubbing on central error logging platforms and pull plaintext credentials out of the logs, which are usually less tightly locked down than the rest of a network. They work through code repositories at machine speed, including commit history that people assume is long gone, looking for credentials and for knowledge of the environment. They spin up their own workloads in Kubernetes and elsewhere, then when hundreds of alerts fire at once defenders have to work out which exposed secrets are still live and rotate them before they're used.

The doors haven't changed
The NCSC reached a similar conclusion in its Annual Review earlier this month, saying the bulk of what it has seen in recent state-backed incidents
"is not an advanced, autonomous end-to-end AI attack, but a much simpler exploitation of common vulnerabilities and exposures (CVEs) and other open vulnerabilities"
For large organisations, 88% of the incidents it handled came in through just four routes:
- exploiting public-facing applications
- phishing
- misuse of valid accounts
- exploiting external remote services
Its forecast is just as plain, saying AI will keep increasing
"the speed, scale and efficiency of existing attack methodologies rather than autonomously conducting end-to-end attacks."
The Bank for International Settlements looked at one of the clearest public examples in its September paper on frontier AI and the financial sector. Between late December 2025 and mid-February 2026 a single attacker used Claude Code and GPT-4.1 to break into nine Mexican government agencies and take more than 150 gigabytes of data. The AI tools sped up exploitation, automated reconnaissance and let one person get through work that would normally need a team of analysts. Claude refused or questioned some requests along the way and the attacker found ways round them. The paper's conclusion was that many of the weaknesses used
"could have been addressed through standard security controls such as patching, credential rotation, network segmentation and endpoint detection."
The clock has changed

CrowdStrike's 2026 Global Threat Report puts the average time for criminal attackers to move on from the first machine they compromise at 29 minutes in 2025, 65% faster than the year before. It also reports that attackers using AI increased their operations by 89% over the year.
Verizon's 2026 Data Breach Investigations Report found that 31% of breaches now start with a software vulnerability, which has overtaken stolen passwords as the most common way in. The BIS paper draws out two more figures from the same report. Stolen credentials accounted for 13% of initial access and only 26% of the critical vulnerabilities on CISA's known-exploited list were fully fixed by organisations in 2025, down from 38% the year before.
The supply of new vulnerabilities is about to grow as well. The BIS paper notes that Anthropic says its Claude Mythos Preview model found more than 10,000 high and critical severity vulnerabilities in some of the world's most important software. More than 99% of them hadn't yet been patched when Anthropic reported. The NCSC expects defenders to face "a race to shorten patch deployment timelines" before attackers can build working exploits.
Google's Threat Intelligence Group put numbers on this at the end of September, reporting that the vulnerabilities disclosed each month doubled this year from 5,045 in January to 10,740 in August and that half of those it could trace to AI led to remote code execution, compared with about a quarter across all disclosures.
"AI is measurably changing not just the pace of vulnerability discovery and exploitation, but also the types and typical risk profiles of vulnerabilities that are being discovered."
Google also found that only around 1 in 431 of this year's disclosed vulnerabilities has been seen exploited, which is why knowing which ones matter in your own environment counts for more than trying to patch everything at once.
There's already a clear example from last month, when on 21 October Wiz disclosed two flaws in PHP, which runs around 70% of the world's websites. It said every PHP version of the last decade is affected by at least one of them. Wiz used Gemini and GPT models throughout the research and said frontier AI had substantially lowered the barrier to exploiting them. The flaws can let an unauthenticated attacker run code or read server memory holding passwords, tokens and API keys. New Scientist reported that the PHP Foundation had seen signs of the problem in 2024 but didn't act, partly because nobody had a practical way to exploit it.
So the problem is less about what attackers can do and more about how long you leave the door open. A vulnerability that used to sit unnoticed for weeks can now be found and used before the next change window comes round.
What the regulators expect
Financial regulators have a clear view on this, with the BIS paper saying financial authorities are converging on a pragmatic response. Rather than writing new AI-specific cyber rules, they are reinforcing the cyber risk and operational resilience frameworks they already have, with more emphasis on quick decision-making, faster patching and better response and recovery. Its conclusion is that frontier AI
"does not fundamentally change the foundations of cyber resilience, but it significantly increases the speed and intensity with which established practices need to be executed."
The paper also points to an argument from the Institute of International Finance that firms will need more tolerance for planned downtime to patch outside the usual maintenance windows. It says this should be seen as responsible security rather than operational failure. That is a change in attitude many change boards haven't made yet.
Credentials last too long
Max's themes are mostly about secrets and the numbers there aren't comforting either. GitGuardian counted 28.65 million new secrets committed to public GitHub repositories in 2025, a 34% rise on the year before. When it retested secrets it had confirmed as valid in 2022, more than 64% of them still worked in January 2026.

That is the part an agent makes worse, because a person going through years of commit history or error logs gets bored and gives up but an agent doesn't. A key that has been valid for four years is worth the same to it whether it leaked last week or in 2022.
What I'd recommend
- Find out how long a leaked credential would stay usable in your environment. If the honest answer is months or years, short-lived credentials and just in time access take most of the value out of anything an attacker finds. NIST's new design tenets for AII agent identity, published on 27 October, say the same for agents and tell organisations to avoid long-lived bearer tokens and static, reusable secrets.
- Check what your error logging platforms actually hold and whether you'd get an alert if someone switched off the scrubbing.
- Scan repositories and their history for secrets and rotate whatever you find rather than just deleting the line.
- Measure how long it really takes to patch a critical vulnerability on an internet-facing system, from the advisory to the fix being live.
- Look for new workloads in your cloud accounts that nobody asked for.
None of this needs an AI security product, as it's the same work we've been asked to do for years, done faster and checked more often.
How long would a leaked key stay usable where you work?
Sources
- Max Henderson, Kroll, LinkedIn post on agentic AI attacks seen in Kroll's cases (6 November 2026): https://www.linkedin.com/feed/update/urn:li:share:7524457102354853889/
- Kroll, Max Henderson profile: https://www.kroll.com/en/our-experts/max-henderson
- NCSC Annual Review 2026 (3 November 2026): https://www.ncsc.gov.uk/collection/annual-review-2026
- Bank for International Settlements, FSI Occasional Paper No 28, When machines attack: frontier AI cyber threats and policy responses in the financial sector (September 2026): https://www.bis.org/publications/fsi-paper-28-when-machines-attack-frontier-ai-cyber-threats-and-policy-responses-financial-sector.pdf
- CrowdStrike, 2026 Global Threat Report press release (24 February 2026): https://www.crowdstrike.com/en-us/press-releases/2026-crowdstrike-global-threat-report/
- Verizon, 2026 Data Breach Investigations Report: https://www.verizon.com/business/resources/reports/dbir/ (13% and the CISA figures as reported in the BIS paper above)
- Google Threat Intelligence Group, Vulnerability Discovery and Exploitation Trends in the AI Era (30 September 2026): https://cloud.google.com/blog/topics/threat-intelligence/vulnerability-discovery-and-exploitation-trends-in-the-ai-era
- GitGuardian, The State of Secrets Sprawl 2026 (17 March 2026): https://blog.gitguardian.com/the-state-of-secrets-sprawl-2026/
- Wiz, PHP2Shell: AI uncovers critical RCE vulnerabilities affecting 70% of the Internet (21 October 2026): https://www.wiz.io/blog/php2shell-php-rce
- New Scientist, Global race to update computers as AI discovers most disruptive hack yet (22 October 2026): https://www.newscientist.com/article/2595282-global-race-to-update-computers-as-ai-discovers-most-disruptive-hack-yet/
- NIST, Design Tenets for Agent Identity Management Systems (27 October 2026): https://www.nist.gov/blogs/cybersecurity-insights/design-tenets-agent-identity-management-systems
