Anthropic revealed in April that its Mythos preview model identified thousands of zero-day vulnerabilities in major operating systems and web browsers, including a 27-year-old bug in OpenBSD. The disclosure has intensified debate across the cybersecurity community, but for Australian government agencies, it sharpens a more urgent problem: AI is compressing the time attackers need to find and exploit weaknesses, and public sector defences are not keeping pace.
Government agencies face a different threat profile from most private sector organisations. They are high-value targets for nation-state actors and sophisticated adversaries with the resources to develop or acquire advanced AI-assisted attack tools. The Australian Signals Directorate's Australian Cyber Security Centre has repeatedly warned that state-backed actors continue to target Australian government networks and critical infrastructure.
With AI, attackers are finding vulnerabilities faster, and the time between discovery and exploitation is shrinking. That places greater pressure on agencies to identify and address risks before they can be used against them.
The gap between AI-enabled attackers and public sector defences
The early assumption was that AI would be a double-edged sword, helping attackers find vulnerabilities but also helping defenders close them. That framing now looks dangerously optimistic. AI is compressing attack timelines faster than most defence organisations can respond, and the gap is widest in the public sector.
Consider the baseline before AI entered the picture: the median time for an organisation to remediate half of its open, internet-facing vulnerabilities was 361 days. Exploitation, by contrast, takes hours. One-third of exploited CVEs in the first half of 2025 showed attacker activity on or before the day of public disclosure, before most teams even knew there was anything to patch.
For federal agencies and state and local government organisations, legacy systems, complex procurement cycles, stretched security staffs, and compliance-heavy environments compound to slow remediation to a near halt. The evidence now points in one direction: in the near term, AI is likely to benefit attackers more than defenders. For the public sector, that imbalance is the problem to solve.
What government agencies can do now
Closing the gap requires action on two fronts: hardening the software development pipeline so vulnerabilities never ship, and increasing proactive threat hunting so attackers don't find exploitable gaps first. Craig Nielsen, Vice President, Asia Pacific & Japan at GitLab, argues that the most underutilised asset in public sector security is proactive offence.
"Red team exercises, bug bounties, and security hackathons are no longer a nice-to-have," Nielsen said in a commentary piece. "If AI is helping attackers discover vulnerabilities at scale, defenders need structured programs to find them first." Agencies that invest in red teaming now will be better positioned than those waiting for CVE disclosures.
Security also needs to shift left, all the way to the developer's integrated development environment. Too many agencies still treat security as a gate at the end of the development process. Hardcoded secrets, known-vulnerable imports, and deprecated API patterns should be flagged before a developer pushes a commit, not during an audit six months later. This approach maps directly to CISA's Secure by Design principles.
For professionals in government roles, building skills in this area is a concrete step. An AI for Cybersecurity Analysts learning path can help teams understand how to deploy AI for vulnerability discovery and remediation within established governance processes.
Enforcement should be consistent across every pipeline. In large agencies with hundreds of active projects, inconsistent enforcement is the rule rather than the exception. Security policies need to be defined once and enforced everywhere, at the merge request level, across every group and project, with exceptions reviewed and logged. If a line of code can enter production without passing a defined set of controls, the gap will be found.
Responding to AI-assisted attacks requires AI-assisted defence. That means deploying security agents that can scan dependency graphs across every project with full context, identify reachable vulnerable call paths, and propose fixes within the same governance process as human-authored changes. Every AI-generated fix must move through the same approval and audit trail as any other change.
Finally, agencies should build for auditability from the start. When an auditor or oversight body asks for evidence that a specific security policy was enforced on a specific change, the answer should be immediately accessible. Evidence generated in flight, such as scan results, applied policies, and merge timestamps, serves as both a compliance and security asset.
For public sector teams looking to build these capabilities internally, AI for Government training resources can support structured upskilling for security and development teams.
Why this matters for government professionals
Anthropic restricted access to its Mythos model, but its own offensive security researchers estimate that comparable attacker tooling will be widely available within six to twelve months. That timeline is short in procurement and delivery terms, but long enough for agencies to act. Strengthening software supply chains, enforcing consistent policy across pipelines, and building proactive threat-hunting capability will reduce exposure while conditions are still shifting. Agencies that move now will be retired substantially better positioned those who wait for the next major exploit.
Your membership also unlocks:
