SafetyRisk, alignment & guardrails
Germany's BSI raises national cyber threat level to 'tense to critical', citing AI-assisted attacks
Germany's federal cyber agency has moved its threat assessment to its highest setting since 2023, saying capable and agentic AI lets attackers find and exploit software flaws faster and cheaper, with average time from disclosure to first attack falling to eight days.
From the Financial Times · AI article of November 19, 2026 and cited sources: The BSI's 2026 report, covering July 2025 to June 2026, names AI as the main reason for the higher threat level. [2] [3] [4]
SOURCES[2] Reuters via Yahoo: German authorities lift threat level to… yahoo.com[3] heise: BSI raises threat level to "tense to critical" heise.de[4] Poland Insight: Germany BSI: Vulnerability Attacks Rise 48% polandinsight.com
AI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- The BSI's 2026 report, covering July 2025 to June 2026, names AI as the main reason for the higher threat level. [2] [3] [4]
- Original source
- Financial Times · AI article, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- November 19, 2026 article published
- Covered
- Nov 20 edition
- Evidence
- 8 sources across 8 sites
- 1Germany hit by AI-driven surge in cyber attacksft.com
- 2Reuters via Yahoo: German authorities lift threat level to…yahoo.com
- 3heise: BSI raises threat level to "tense to critical"heise.de
- 4Poland Insight: Germany BSI: Vulnerability Attacks Rise 48%polandinsight.com
- 5TextilWirtschaft: Claude Mythos findet…textilwirtschaft.de
- 6AISI: Our evaluation of Claude Mythos Preview's cyber…aisi.gov.uk
Reach: not measured yet
The BSI's 2026 report, covering July 2025 to June 2026, names AI as the main reason for the higher threat level. Its figures show exploitation attempts up 48% and vulnerability scanning up 57%. The average gap between a flaw becoming public and the first attack on it fell from about 53 days two years ago to 8 days in Q2 2026. The agency also sees attackers moving away from email-based attacks and going after systems directly. [2] [3] [4]
The most concrete public evidence of AI attack capability comes from the UK AI Security Institute's test of Anthropic's unreleased Claude Mythos Preview. The model solved 73% of expert-level capture-the-flag challenges. It also completed a full 32-step simulated corporate network attack in 3 of 10 runs. The test networks had no active defenders, endpoint detection or incident response. [6] [5]
Midsize firms overrate their own security. About 90% describe it as good or very good, yet only 56% meet basic requirements in the BSI's Cyber Risk Check. Small and medium-sized firms account for 90% of ransomware cases. The BSI also says companies have been slow to use AI to scan their own systems for flaws. [3] [4] [2]
The BSI's figures show faster, more frequent exploitation; AI's causal share is the agency's assessment, not case-by-case attribution. If the trend holds, monthly patch cycles and self-rated security look inadequate for internet-facing systems.
The sources are gathered; the dossier follows.
Executive brief
Germany's federal cyber agency, the BSI, has raised its national IT threat assessment to "tense to critical", its highest setting since 2023. It names AI as the main reason. Capable AI models, including agentic systems, help attackers find and exploit software flaws faster, more cheaply and with more precision (Reuters via Yahoo). The gap between a flaw becoming public and the first attack on it averaged 8 days in Q2 2026. Two years ago it was about 53 days (heise). The FT headline also uses the word "critical". Only the FT's teaser is available, not the full article.
What changed and event timeline
React2Shell exploited within hours
CVE-2025-55182 carries the maximum severity score (CVSS 10). Attacks began within hours of a public proof-of-concept. The BSI uses it to show how short the window for patching has become ().
BSI weighs in on Claude Mythos
BSI President Claudia Plattner said the agency is talking with Anthropic about Mythos and expects "upheavals" in how vulnerabilities are handled. The BSI had not tested the model itself ().
UK government institute tests Mythos Preview
The UK AI Security Institute published results showing Mythos Preview completed a 32-step simulated network attack from start to finish in 3 of 10 runs ().
BSI publishes its 2026 report
The report covers July 2025 to June 2026. Attempts to exploit vulnerabilities rose 48% and scanning for weak spots rose 57% (;).
Threat level raised
The BSI says email-based attacks are falling as attackers use AI to go after systems directly. It says companies are not strengthening their defences fast enough ().
Capabilities and access
- The BSI report describes "high-performance" and agentic AI in general terms.
- The model most clearly linked to the BSI's concern is Anthropic's Claude Mythos Preview. It is not publicly released.
- OpenAI responded with a security product called "Daybreak" (Börse Express).
Read the full section
- The BSI report describes "high-performance" and agentic AI in general terms. Coverage of the report does not tie the new threat level to any named model (heise).
- The model most clearly linked to the BSI's concern is Anthropic's Claude Mythos Preview. It is not publicly released. Access goes through partnerships with Apple, Amazon, Microsoft, CrowdStrike, Palo Alto Networks and others (TextilWirtschaft).
- OpenAI responded with a security product called "Daybreak" (Börse Express).
Technical analysis for researchers and developers
- How the main independent test worked (AISI):
- Capture-the-flag (CTF) hacking challenges at several difficulty levels. Mythos Preview solved 73% of the expert-level ones.
- "The Last Ones", a 32-step attack on a simulated corporate network that takes a human about 20 hours.
Read the full section
- How the main independent test worked (AISI):
- Capture-the-flag (CTF) hacking challenges at several difficulty levels. Mythos Preview solved 73% of the expert-level ones.
- "The Last Ones", a 32-step attack on a simulated corporate network that takes a human about 20 hours. Mythos Preview averaged 22 steps; Claude Opus 4.6 averaged 16.
- Where the test falls short: the simulated networks had no active defenders, no endpoint detection and no incident response. AISI says it cannot tell whether the model could break into well-defended systems.
- What the BSI measures: its figures are counts of national activity (exploit attempts, scans, time to exploitation). They are not attributed to AI one case at a time, so they show a trend that fits AI use but do not measure AI's share.
Claims and evidence
- Threat level raised; AI is the driver: this is the BSI's own assessment (Reuters via Yahoo).
- Reuters reports 238 points in mid-2026, up from 124 a year earlier.
- 35% more new vulnerabilities, about 160 published per day: BSI data (heise).
Read the full section
- Threat level raised; AI is the driver: this is the BSI's own assessment (Reuters via Yahoo).
- The BSI Cyber Index figures conflict. Reuters reports 238 points in mid-2026, up from 124 a year earlier. Poland Insight reports 170, about 50% above the previous period, with 2024 set at 100 (Reuters via Yahoo; Poland Insight). The primary BSI document was not among the reviewed sources, so it is unclear which figure is right.
- 35% more new vulnerabilities, about 160 published per day: BSI data (heise).
- "AI-assisted break-in and data theft in 25 minutes": this comes from Unit 42, the threat research team at Palo Alto Networks, a security vendor. It is not a BSI measurement (Börse Express).
- Anthropic says Mythos found "thousands" of serious vulnerabilities: this is reported in the press, not independently checked (TextilWirtschaft).
Context and prior work
- In its 2025 report the BSI rated the situation "tense" and named phishing as the main way attackers get in (Reuters via Yahoo).
- Germany recorded 1,041 ransomware cases in 2025, up from 950 the year before (Börse Express).
- Reported AISI figures put the doubling time for frontier AI's offensive cyber capability at about four months (Decrypt).
Read the full section
- In its 2025 report the BSI rated the situation "tense" and named phishing as the main way attackers get in (Reuters via Yahoo). The 2026 report describes a shift away from email-based attacks.
- Germany recorded 1,041 ransomware cases in 2025, up from 950 the year before (Börse Express).
- Reported AISI figures put the doubling time for frontier AI's offensive cyber capability at about four months (Decrypt).
Limitations, safety and contested findings
- The two Cyber Index figures above cannot both be right.
- The strongest evidence for AI attack capability comes from test networks with no defenders.
- The national statistics cover all attacker activity. In the reviewed sources, AI's causal role rests on the BSI's assessment rather than on attack-by-attack attribution.
Read the full section
- The two Cyber Index figures above cannot both be right.
- The strongest evidence for AI attack capability comes from test networks with no defenders. AISI itself says it does not show success against hardened systems (AISI).
- The national statistics cover all attacker activity. In the reviewed sources, AI's causal role rests on the BSI's assessment rather than on attack-by-attack attribution.
- Plattner has criticised leading AI companies for paying too little attention to the security of their own systems (CIO).
Business and practitioner implications
- Patching: with an average of 8 days between disclosure and first attack, monthly patch cycles are too slow for systems exposed to the internet (heise).
- Self-assessment gap:
- 90% of midsize firms rate their security "good" or "very good", but only 56% meet basic requirements in the BSI's Cyber Risk Check (heise).
Read the full section
- Patching: with an average of 8 days between disclosure and first attack, monthly patch cycles are too slow for systems exposed to the internet (heise).
- Self-assessment gap:
- 90% of midsize firms rate their security "good" or "very good", but only 56% meet basic requirements in the BSI's Cyber Risk Check (heise).
- 90% of ransomware cases hit small and medium-sized firms (Poland Insight).
- Use AI defensively: the BSI says firms have been slow to use AI to scan their own systems for flaws before attackers do (Reuters via Yahoo).
- Regulation: expect the EU Cyber Resilience Act to set minimum security standards for devices. The BSI also recommends building new systems so encryption methods can be swapped later ("crypto-agility") (heise).
The source trail.
Sources (8)
Germany hit by AI-driven surge in cyber attacks
Feed excerpt only.
ft.com