HLD Pulse sector briefingFinancial servicesCritical infrastructure

Defending the financial system

How HLD Sentinel protects banks, insurers, super funds and the infrastructure that moves money.

Financial services took more of Australia's critical infrastructure incidents last year than any other sector. This briefing sets out how attackers are reaching financial institutions in 2026, why the shared infrastructure beneath them raises the stakes, and how Sentinel detects and contains those attacks inside limits a regulated entity can defend to its board and to APRA.

HLD Shield Threat Intelligence12 minute read
of critical infrastructure incidents ASD handled in 2025-26 were in financial and insurance services
30%
of critical infrastructure incidents ASD handled in 2025-26 were in financial and insurance services
ASD Annual Cyber Threat Report 2025-26
average cost of a data breach in financial services, 26% above the global average
US$6.29M
average cost of a data breach in financial services, 26% above the global average
IBM Cost of a Data Breach 2026
of breaches involved a third party such as a supplier or service provider
48%
of breaches involved a third party such as a supplier or service provider
Verizon DBIR 2026
of breaches began with an exploited vulnerability, up from 20% a year earlier
31%
of breaches began with an exploited vulnerability, up from 20% a year earlier
Verizon DBIR 2026

01

Why finance is targeted

Every financial institution holds two things attackers want: money that can be moved, and the data that proves who owns it. A bank, insurer or super fund keeps both in volume, behind systems that have to stay available around the clock. That combination draws every kind of attacker at once. Fraud crews want access to the payment rails. Ransomware operators want the outage, because a business that cannot settle payments for a day is under heavy pressure to pay. State actors want the data and, in a crisis, the ability to disrupt.

The Australian figures bear this out. The Australian Signals Directorate's Annual Cyber Threat Report 2025-26, released this month, records 1,524 incidents handled by the Australian Cyber Security Centre, up 20% on the previous year. Financial and insurance services accounted for 30% of the critical infrastructure incidents in that total, ahead of public administration and safety at 21% and the energy, water and waste utilities at 17%. Where the initial access was known, phishing was the most common way in, at 36%.

Getting it wrong also costs more in finance than almost anywhere else. IBM's 2026 Cost of a Data Breach report puts the average breach in financial services at US$6.29 million, about 26% above the global average of US$4.99 million. Only healthcare costs more. Those figures leave out the losses that never appear on an incident invoice: customers who move their accounts, remediation programs agreed with a regulator that run for years, and the board time spent explaining what happened.

02

What October showed

October 2026 has been called Black October in the security press. The name describes a run of unrelated disclosures and attacks that landed within days of each other, and several of them touched finance directly.

On 21 October the State Bank of India told the stock exchange it had "experienced an incident involving the loss of certain data", after CERT-In and the Defence Cyber Agency flagged a dataset for sale on the dark web. The bank has confirmed 3,897 of its customer records inside a dump of around 375,000, exposing names, phone numbers, home addresses, dates of birth and loan details. The same day, Japan's Onamae Desktop Cloud began shutting down customer environments after intruders installed malware on its infrastructure, with more than 26,000 servers potentially affected. ScruteX, which tracked 28 incidents across the month, describes many of those servers as FX trading machines.

Other incidents hit the software financial firms build on. The PHP2Shell flaws (CVE-2026-109618 and CVE-2026-109619) allowed unauthenticated remote code execution on PHP servers and went onto CISA's Known Exploited Vulnerabilities catalog. On 21 October a compromised release of LiteLLM, an open-source gateway many teams use to route requests to AI models, sat on PyPI and Docker Hub for a few hours. Any environment that pulled it may have handed over the API keys and cloud credentials it held. Of ScruteX's 28 October incidents, 20 were confirmed in use in real attacks, and ten surfaced within a single 48-hour window on 20 and 21 October.

Most of these could be patched or contained within hours of disclosure. The institutions that suffered were the ones that could not work out quickly whether they were exposed, and then could not act on the answer before an attacker did. Closing that gap between knowing and acting is the subject of the rest of this briefing.

03

How attackers are getting in

Across the incidents HLD has tracked this year, attackers keep reaching financial institutions by the same five routes.

Identity and the help desk

The cheapest way into a bank is still to persuade someone to let you in. ASD's report describes an attempted voice phishing attack on an Australian bank in August 2025, in which two callers impersonating a Melbourne employee rang the internal help desk and asked about staff in a specialised financial services team. In March and April 2025, credential stuffing attacks on AustralianSuper, Australian Retirement Trust, Rest, Hostplus and Insignia Financial used passwords leaked in other breaches to log into member accounts; AustralianSuper said up to 600 stolen passwords were used. None of these attacks needed malware. Each produced sign-ins that looked legitimate in isolation and became suspicious only when someone connected them to what happened next.

Suppliers and shared platforms

Verizon's 2026 Data Breach Investigations Report found a third party involved in 48% of breaches. Finance is especially exposed because so many institutions run on the same small set of core banking, payments, hosting and managed service providers. In July 2024 a ransomware attack on C-Edge Technologies, which ran payment systems for around 300 small co-operative and regional banks in India, forced those banks off the national payments network while it was contained. Later that month a faulty CrowdStrike update took down Windows machines worldwide, including at banks, payment processors and exchanges. That second event was an accident, yet for customers the effect was much the same, and it is exactly the kind of disruption APRA's CPS 230 now expects regulated entities to plan for.

Exposed software at the edge

Exploited vulnerabilities were the way in for 31% of breaches in Verizon's 2026 data, up from 20% a year earlier. VPN gateways, file transfer servers, customer portals and the web applications behind them sit on the internet by design. When a flaw like PHP2Shell or this year's FortiOS SSL-VPN overflow is published, mass scanning starts within hours. A patch cycle measured in weeks leaves an institution open for most of that time.

Ransomware and destructive malware

Ransomware appeared in 48% of the breaches Verizon analysed this year. In finance the encryption is usually the second act: operators take data first, then encrypt or wipe systems to force a decision. Tools like GigaWiper, which HLD covered in July, go further by throwing away their own encryption keys, so restoring from backup becomes the only way back. For a payments business the real question during a ransomware event is how long critical services can stay down before customers and the market notice, and the honest answer is usually a matter of hours.

Machine identities and AI agents

The newest route runs through credentials that belong to software. API keys, service accounts, deployment pipeline tokens and, increasingly, AI agents with access to customer data or internal tools all authenticate with no person present. The LiteLLM compromise showed how fast one poisoned dependency can collect those keys at scale. Many institutions cannot list every machine identity able to move money or read customer records, so they have no way to tell when one of them starts behaving differently.

04

The infrastructure underneath raises the stakes

Individual banks are the visible part of the system. Beneath them sit payment switches, clearing and settlement systems, card networks, market infrastructure, and the cloud and data centre providers that host much of it. These shared layers are where a single compromise becomes a sector-wide event.

When a shared platform goes down, every institution connected to it goes down with it, whatever the quality of their own controls.

Australian law already treats this layer as critical. The Security of Critical Infrastructure Act 2018 covers banking, superannuation, insurance, financial market infrastructure and payment systems, and the entities responsible for those assets must report significant cyber incidents to government, within 12 hours where the impact is significant. APRA's CPS 230 Operational Risk Management standard, in force since 1 July 2025, requires regulated entities to identify their critical operations, set tolerance levels for disruption and manage the material service providers those operations depend on. From 1 July 2026 it reaches existing supplier contracts as well. CPS 234 continues to require information security capability that matches the threat, controls that are tested for effectiveness, and notice to APRA within 72 hours of a material incident.

Regulators are also checking. APRA's tripartite cyber assessments now cover more than 300 banks, insurers and super trustees, and the early rounds found significant control gaps. ASIC called on financial firms in May 2026 to lift their cyber resilience, and in July the government told the major superannuation associations to do the same. Across all of this, regulators are asking for evidence that controls work in practice, and a policy document on its own no longer satisfies them.

05

How Sentinel defends a financial institution

HLD Sentinel is built for organisations that face serious threats without a large security operations team. Outside the major banks, that describes most of the financial sector: mutual banks and credit unions, insurers, super funds, brokers, fintechs, payment providers and the service companies that support them. Sentinel deploys itself into an environment, defends it within boundaries the organisation authorises, and keeps checking that its own protection still works.

Identity, email, network and cloud seen together

Almost every attack described above leaves traces in more than one system. A help-desk password reset is followed by a sign-in from an unfamiliar device, which is followed by a mailbox rule hiding replies from a payments supplier. Sentinel watches identity, email, network and cloud as one surface and correlates those events, so the analyst sees a single incident with its full sequence where they would otherwise get three low-priority alerts in three consoles. Human, machine and AI-agent identities sit in the same security model, which matters when the account doing something unusual is a service account or an agent with access to customer records.

Detection based on behaviour as well as signatures

Signature rules catch what has been seen before. Sentinel adds behavioural baselines and cross-source correlation, so it can flag activity that is unusual for a particular user, device or service even when no rule exists for it yet. That is the capability that matters on the day a new flaw is published and nobody has written a signature for it.

Response that stays inside your limits

Speed is of little use to a regulated entity if it cannot explain what its software did. Sentinel's response is graded. Low-risk actions such as ending a session or rotating a compromised password happen automatically. Heavier actions wait for human approval, within limits the institution sets. Every action is reversible and recorded with the evidence that prompted it. The organisation sets a hard cap on how many actions Sentinel may take in an hour, administrator accounts are excluded from automatic action by default, and a kill switch halts all enforcement immediately. For an institution whose critical operations carry tolerance levels measured in hours, those guardrails are what make automated response acceptable at all.

Protection that checks itself

Control testing often finds that a security tool was set up correctly once and has degraded since. A log source stopped sending, an agent was removed during a rebuild, or a new cloud account went live without monitoring. Sentinel continuously looks for broken integrations, telemetry gaps, configuration drift and weakened protection, and raises each one as a finding. CPS 234 asks institutions to test that their controls are effective; a gap found and fixed this week is far easier to defend than one an auditor discovers next year.

Containment on the endpoint

On staff laptops and desktops, a lightweight Sentinel agent collects process and network telemetry, detects tampering and can isolate a compromised machine without waiting for someone to answer the phone. That shortens the window between an infostealer landing on a staff laptop and the credentials it collected being used against the payments platform.

Published evidence

Against a fixed-signature control built on Sigma rule conventions, with scenarios classified using MITRE ATT&CK and both systems given identical inputs, Sentinel measured:

+80%

detection coverage

60%

fewer items for human review

45%

fewer manual routing decisions

Corpus, formulas and raw counts are public. Check the method.

06

A morning at a regional lender

The scenario below is illustrative. It combines patterns from the incidents in this briefing to show how detection, correlation and graded response work together during a real attack.

  1. 08:42

    A convincing call to the service desk

    Attacker

    A caller claiming to be a treasury analyst, locked out before a settlement run, asks the service desk for help. The desk resets the password and enrols a new phone for multi-factor authentication.

  2. 08:51

    A sign-in that only looks routine

    Sentinel

    The account signs in from a residential connection in another state, on a device Sentinel has never seen. On its own this is low risk. Sentinel links it to the reset and the new MFA device nine minutes earlier and raises the risk score.

  3. 08:55

    Automatic containment

    Sentinel

    The session creates an inbox rule that files messages from the lender's payments processor into an archive folder, then starts exporting supplier bank details. Sentinel ends the session and rotates the password, both actions inside its approved hourly limit.

  4. 08:56

    One incident, with the evidence attached

    Sentinel

    The security lead receives a single incident showing the reset, the MFA change, the sign-in, the inbox rule and the export, with each automatic action listed and reversible. Sentinel recommends removing the new MFA device, an action that waits for approval.

  5. 09:04

    Human decision, recorded

    Security lead

    The security lead approves the removal, confirms with the real analyst by phone and closes the containment stage. The incident record holds what the risk team needs to decide whether the event is notifiable under CPS 234.

Without correlation, the same morning produces a routine password reset, a sign-in alert that someone triages after lunch and a supplier payment redirected at the end of the week.

07

Mapping Sentinel to your obligations

Boards and risk committees will ask how a security platform supports the institution's regulatory position. The table sets out where Sentinel contributes evidence against the main Australian requirements.

FrameworkRequirementWhere Sentinel contributes
APRA CPS 234Information security capability that matches the threats the entity facesContinuous detection across identity, email, network, cloud and endpoints, measured against a published benchmark.
APRA CPS 234Systematic testing of control effectivenessSelf-monitoring raises telemetry gaps, broken integrations and configuration drift as findings, with dates.
APRA CPS 234Notify APRA within 72 hours of a material information security incidentCorrelated incident records with a timeline and the evidence behind each action, ready for the materiality assessment.
APRA CPS 230Keep critical operations within tolerance levels during disruptionAutomatic containment of low-risk actions shortens time to contain, while approval gates keep critical systems from being disrupted without sign-off.
APRA CPS 230Manage risk from material service providersMonitoring of supplier connections, third-party access and machine identities for behaviour that departs from the baseline.
SOCI Act 2018Report significant cyber incidents to government within 12 hoursTimestamped detection and response records that support a fast, accurate first report.

Sentinel supports these obligations. It does not by itself make an institution compliant, and each entity remains responsible for its own assessment and reporting decisions.

08

What to do this quarter

These steps are worth taking whether or not an institution uses Sentinel. Each one closes a gap that appeared in at least one incident in this briefing.

  1. 01

    Test the help desk against a voice phishing call

    Run an authorised exercise in which someone impersonates a staff member and asks for a password reset and a new MFA device. For any account with payment or treasury access, require a call-back to a number already on file before a reset goes ahead.

  2. 02

    Map your machine identities

    List the service accounts, API keys and AI agents that can move money or read customer data, who owns each one and when each was last rotated. Rotate anything that may have been exposed during the LiteLLM or PHP2Shell windows.

  3. 03

    Compare patch time with exploitation time

    For internet-facing systems, compare how long patching took for the last three critical advisories with how long attackers took to start scanning. Where patching was slower, move that class of system to emergency change.

  4. 04

    Prove your telemetry

    Pick ten critical systems and confirm their logs are reaching your monitoring this week. Many control gaps begin with a log source that stopped sending months ago and nobody noticed.

  5. 05

    Rehearse a supplier failure

    Choose one material service provider under CPS 230 and walk through a day in which it is compromised or unavailable. Record how long each critical operation stays within its tolerance level.

  6. 06

    Agree automation limits before you need them

    Decide with the board which actions security tooling may take without approval, which need sign-off and who can give it. Settling this in advance is far easier than arguing about it during an incident.

09

Closing assessment

The financial system depends on confidence that money will be where it is supposed to be. In 2026 attackers have more ways to undermine that confidence than they did a year ago, and they move faster. Exploitation begins within hours of disclosure, credentials are harvested through the tools developers trust, and social engineering targets the people whose job is to help. Defenders in the sector cannot match that pace by hiring more analysts, and most of them could not find the analysts if they tried. What they can do is connect events across systems, contain attacks in minutes inside limits they have approved, and show on any given day that their controls are working.

Sentinel is built for exactly that job. HLD is onboarding financial institutions into the Sentinel public beta in small groups, each with a scoped evaluation agreed in advance: the systems in scope, the baseline Sentinel is compared against and the results that would count as success.

Evaluate Sentinel against your own environment

Tell us which systems matter most and what success would look like. We will agree the scope, the baseline and the acceptance criteria before the evaluation starts.

More HLD Pulse briefings

Incident analysis and sector intelligence from HLD Shield.

Back to HLD Pulse