SafetyRisk, alignment & guardrails
Australia's cyber agency tells organisations to put security ahead of uptime, warning online services may go down without notice
Australia's signals intelligence agency says organisations should prioritise security over availability and warns that essential online services may go offline without notice. Its annual threat report cites rising incident volumes and faster AI-assisted attackers, and urges defenders to adopt frontier AI tools.
From the The Guardian · AI article of October 26, 2026 and cited sources: The ASD's 2025–26 cyber threat report says organisations should accept downtime to secure their systems. [1] [9]
SOURCES[1] Australians should prepare for online services to go down… theguardian.com[9] NewsBytes: Australia's cyber agency urges smarter AI tools newsbytesapp.com
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- The ASD's 2025–26 cyber threat report says organisations should accept downtime to secure their systems. [1] [9]
- Original source
- The Guardian · AI article, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- October 26, 2026 article published
- Covered
- Oct 26 edition
- Evidence
- 13 sources across 13 sites
- 1Australians should prepare for online services to go down…theguardian.com
- 2Anthropic: Project Glasswinganthropic.com
- 3ACS Information Age: Anthropic gives Mythos AI to…ia.acs.org.au
- 4ABC News: Australia needs an AI "early warning system"abc.net.au
- 5watchTowr: PHP2Shell FAQwatchtowr.com
- 6Tech Champion: Inside the PHP emergency patchestech-champion.com
Reach: not measured yet
The ASD's 2025–26 cyber threat report says organisations should accept downtime to secure their systems. The agency responded to about 1,500 incidents, up 20%. It sent more than 4,400 notifications about possible malicious activity, up 80%. These figures come from the agency and have not been independently audited. [1] [9]
The ASD says AI is making attackers faster and more effective, and it wants defenders to use frontier AI to find weaknesses, detect suspicious behaviour and respond faster. The agency already uses Anthropic's restricted Claude Mythos Preview through Project Glasswing. That access is subject to US export controls, and Australian organisations briefly lost it in June. [1] [3] [4] [7]
Days before the report, two critical PHP remote-code-execution flaws, known as php2shell, leaked ahead of their fix. Emergency patching took government websites offline in several countries. Exposure was highest for Apache mod_php and for unsupported PHP branches. This matches the ASD's warning that legacy systems are the most attractive targets. [5] [6] [8]
The evidence: the ASD reports more incidents and notifications, and October's emergency PHP patching already took government sites offline.
Read the full assessment
The likely implication is that continuity plans, customer messaging and patch processes should assume deliberate downtime, and that AI-assisted defence depends on access controlled by foreign export policy.
Sources checked; writing the dossier now.
Executive brief
Australia's signals intelligence agency now advises organisations to put security ahead of uptime. The warning comes in the Australian Signals Directorate's (ASD) 2025–26 annual cyber threat report. ASD responded to 1,500 incidents that year, up 20%.
Read the full section
Australia's signals intelligence agency now advises organisations to put security ahead of uptime. In practice, that means Australians should expect more important online services to go offline without warning while they are secured (Guardian). The warning comes in the Australian Signals Directorate's (ASD) 2025–26 annual cyber threat report. ASD responded to 1,500 incidents that year, up 20%. Its notices to outside organisations about possible malicious activity rose 80% to more than 4,400. The agency says AI is making attackers faster and more effective, and wants defenders to use frontier AI themselves. ASD already has access to Anthropic's restricted Claude Mythos Preview model. Days before the report, emergency PHP patches had already forced government websites offline in several countries.
What changed and event timeline
Anthropic launches Mythos Preview under Glasswing
Anthropic gave its vulnerability-hunting model only to selected defenders and committed $100M in usage credits. It did not release the model publicly ().
Australia gains Mythos access
Project Glasswing expanded to 15 more countries, roughly 200 organisations in total. ASD and some Australian companies were included ().
Bradshaw calls for an AI "early warning system"
The ASD chief proposed formal sharing of AI risk information between agencies and AI labs. She also confirmed agency access to Mythos, which is subject to US export controls ().
20–21 Oct 2026: The "php2shell" bugs leak, then get emergency patches
Details of two unauthenticated remote-code-execution flaws in PHP leaked a day before the planned fix. Governments took sites offline while patching (;).
ASD warns of more outages
Bradshaw and Defence Minister Richard Marles said organisations should accept downtime to secure their systems. The Guardian says the report itself was released "on Tuesday" ().
Capabilities and access
- It is not generally available. Anthropic has set post-preview pricing at $25/$125 per million input/output tokens (Anthropic).
- Who can use it: national security agencies, critical-infrastructure defenders and selected companies in partner countries (ACS).
- Access can be withdrawn: Australian organisations lost access to Mythos for about two weeks in June.
Read the full section
- Model: Claude Mythos Preview. It is not generally available. Anthropic has set post-preview pricing at $25/$125 per million input/output tokens (Anthropic).
- Who can use it: national security agencies, critical-infrastructure defenders and selected companies in partner countries (ACS).
- Access can be withdrawn: Australian organisations lost access to Mythos for about two weeks in June. The US kept the right to restrict it again (Startup Daily).
- How ASD uses it: to help Australian organisations find vulnerabilities and harden their systems (ABC).
Technical analysis for researchers and developers
- What ASD has documented: none of its own AI tooling methods.
- php2shell, the vulnerability: one bug is triggered by out-of-memory corruption, the other by stale global state.
- php2shell, the fix: upgrade to PHP 8.5.12, 8.4.27, 8.3.36 or 8.2.35.
Read the full section
- What ASD has documented: none of its own AI tooling methods. The report recommends AI for finding weaknesses, spotting suspicious behaviour and speeding up response (Guardian).
- php2shell, the vulnerability: one bug is triggered by out-of-memory corruption, the other by stale global state. They affect mod_php, PHP-FPM, php-cgi, FrankenPHP and LiteSpeed. Unsupported PHP 5 and 7.x branches are also exposed (watchTowr).
- php2shell, the fix: upgrade to PHP 8.5.12, 8.4.27, 8.3.36 or 8.2.35. Apache mod_php carries the highest risk because an attack can succeed before any application code runs (CISA).
- Interim steps: CISA recommends web application firewalls and SIGMA detection rules. If a system cannot be patched, it should be taken offline.
Claims and evidence
- Incident counts (1,500, +20%; 4,400+ notifications, +80%) come from ASD as reported by the Guardian and repeated by NewsBytes.
- Mythos capability figures come from Anthropic only. The reviewed sources contain no independent evaluation.
- State-actor "pre-positioning" in the Indo-Pacific is ASD's own assessment. It is not corroborated by the reviewed sources.
Read the full section
- Incident counts (1,500, +20%; 4,400+ notifications, +80%) come from ASD as reported by the Guardian and repeated by NewsBytes. These are agency-reported figures that have not been independently audited.
- Mythos capability figures come from Anthropic only. Examples are "thousands" of zero-days and CyberGym 83.1% vs 66.6% for Opus 4.6 (Anthropic). The reviewed sources contain no independent evaluation.
- State-actor "pre-positioning" in the Indo-Pacific is ASD's own assessment. It is not corroborated by the reviewed sources.
Context and prior work
- Last year's report (2024–25) also flagged faster attacks on critical infrastructure and gave boards a list of priorities that included legacy IT (Cyber News Centre; AICD).
- What is new this year: the explicit trade of availability for security, and the formal reliance on access to frontier models.
- Glasswing's starting partners included AWS, Apple, Google, Microsoft, CrowdStrike and the Linux Foundation (Anthropic).
Read the full section
- Last year's report (2024–25) also flagged faster attacks on critical infrastructure and gave boards a list of priorities that included legacy IT (Cyber News Centre; AICD).
- What is new this year: the explicit trade of availability for security, and the formal reliance on access to frontier models.
- Glasswing's starting partners included AWS, Apple, Google, Microsoft, CrowdStrike and the Linux Foundation (Anthropic).
Limitations, safety and contested findings
- Who found the PHP bugs is disputed. Axios headlined them as found with OpenAI technology, and TechBriefly reports the FBI said AI was used.
- Whether the bugs were exploited is also disputed.
- Relying on a foreign-controlled model is a risk. Access to Mythos depends on US export policy (Startup Daily).
Read the full section
- Who found the PHP bugs is disputed. Axios headlined them as found with OpenAI technology, and TechBriefly reports the FBI said AI was used. Meanwhile watchTowr and Tech Champion credit researchers at Calif (and Wiz) and do not mention AI.
- Whether the bugs were exploited is also disputed. At publication, CISA had seen no successful exploitation. watchTowr reports exploitation from 21 Oct, seen in its honeypots.
- Relying on a foreign-controlled model is a risk. Access to Mythos depends on US export policy (Startup Daily).
Business and practitioner implications
- ASD now treats taking a service offline to secure it as acceptable. Continuity plans and customer communications should assume this will happen (Guardian).
- ASD names unsupported systems as the most attractive targets. Old PHP branches are a current example.
- Patching windows are shrinking. With php2shell, attackers moved within hours of the leak.
Read the full section
- Plan for deliberate outages. ASD now treats taking a service offline to secure it as acceptable. Continuity plans and customer communications should assume this will happen (Guardian).
- Fix legacy systems first. ASD names unsupported systems as the most attractive targets. Old PHP branches are a current example.
- Patching windows are shrinking. With php2shell, attackers moved within hours of the leak.
- Board accountability is explicit. ASD calls cyber a "core business risk and leadership responsibility".
- AI-assisted defence depends on vendors. Practitioners should budget for that work while assuming access to frontier models can be cut off.
The source trail.
Sources (13)
Australians should prepare for online services to go down unexpectedly as threats rise, cyber agency warns
Article text retrieved; extracted text may omit tables or interactive elements.
theguardian.com