TL;DR
- Before an intrusion, attackers study the target: staff profiles, job ads, company posts, event photos. AI agents now do this continuously and in parallel.
- Anthropic's September 2026 threat report describes a fleet of 13 scheduled collection agents and concludes that the techniques are old while the economics have changed, making "previously marginal targets viable".
- In a study with real participants, spear-phishing emails researched and written entirely by AI agents got a 54% click rate, level with human experts, against 12% for generic emails.
- For a company, this means the detailed dossier is no longer reserved for the CEO. The accounts clerk and the new helpdesk hire get one too.
- Most of the defence sits in management decisions, what you publish, how money moves, who owns the exposure, and very little of it requires a new tool.
Assumptions
- Sources: Anthropic's threat report, a Trend Micro proof of concept, a peer-reviewed study by Heiding, Lermen, Kao, Schneier and Vishwanath, and Verizon's 2026 DBIR. Vendor reports explain mechanisms well and measure prevalence poorly, since their authors also sell the defence. Solid on mechanism, medium on how widespread it is.
- This piece is about defence and contains nothing operational for attackers.
- No company can become invisible. The aim is to make reconnaissance more expensive and what it finds less useful.
Goal
To explain in business terms what changes when reconnaissance is done by an agent, and to set out five decisions, in order of priority, that a management team can take this quarter.
The part of the attack nobody sees
Every serious intrusion starts with homework. MITRE ATT&CK treats reconnaissance as a tactic of its own: learning about the organisation, its people and its exposed systems, often without touching a single server. Someone reads profiles, cross-references names, works out who approves payments and who joined the finance team last month.
It still works. Verizon's 2026 DBIR puts the human element in 62% of breaches. What kept that homework rare was its cost, because a credible message needs someone who knows who they are writing to.
What an agent changes
Anthropic's report describes reconnaissance that never stops. One loop maps the targets' exposed surface and feeds a project memory that widens the target list on the next pass; in another operation, thirteen standing agents downloaded content from public sites and social profiles on a schedule, summarised it and scored it. The report's own conclusion is the one a board should read: AI "compresses the cost side of attacker ROI calculations", and that shift "makes previously marginal targets viable".
Trend Micro showed the same thing with tools anyone can use, going from a leadership team's public LinkedIn posts to tailored content for each person in about 30 minutes. The Heiding study measured the last step, with the 54% click rate above.
Put those together and the target list changes shape. Deep profiling used to be worth the effort only for the CFO or the system administrator. When an agent does the work, it becomes worth doing for the person who processes supplier invoices, who usually has fewer defences and more reasons to be helpful. (Agents also make plenty of mistakes, including invented credentials and public data passed off as secret. A collector that cannot tell true from false can be tripped up, see point 5.)
Five decisions for management, in order
1. Commission your own reconnaissance, every quarter. Someone, internal or external, collects what an attacker would find in an afternoon: an org chart rebuilt from profiles, the email address format, technologies named in job ads, metadata in published documents, exposed services. The output is a list of things to close, each with an owner and a date.
2. Decide what the company publishes. Job ads can ask for ERP experience without naming version and configuration. Documents go out stripped of metadata. Roles that receive outside requests use function addresses. Out-of-office replies to external senders give a return date and leave out the deputy's name and number. Each detail looks harmless on its own; the value for an attacker comes from combining them, which is exactly what agents are good at.
3. Make out-of-band verification the rule for money and access. Bank detail changes, urgent transfers, password or MFA resets: call back on a number you already hold, never the one in the message, and agree a code word for executive requests. It is the cheapest control on this list and the one with the best return.
4. Use phishing-resistant authentication where it counts. Passkeys or FIDO2 keys for executives, finance, administrators and helpdesk, plus a written procedure for resetting a second factor, because that phone call is where a well-researched attacker does the most damage. A good dossier makes the bait credible; a passkey makes the stolen password useless.
5. Train for the attack you will actually get, and set a few traps. Generic phishing simulations teach people to spot the 12%. Run personalised exercises, make "I'll call you back" a normal answer, and publish a handful of decoy addresses or profiles that only a scraper would reach, wired to an alert. When someone writes to an employee who does not exist, you know the reconnaissance has started.
Why this matters for your business
Trend Micro's thirty minutes cuts both ways. Spend half an hour this week on your own public footprint, the profiles, the job ads, the last press release, and ask what someone trying to fool your accounts team would take from it. Whatever you find goes at the top of the list in point 1.
— Pan
A longer version for security practitioners, in Italian, is on pallanca.net.